Skip to content

Index

ipw.agents.mcp

MCP (Model Context Protocol) server implementations.

This module provides unified interfaces for: - Local models (via Ollama, vLLM) - Cloud APIs (OpenAI, Anthropic, Gemini, OpenRouter) - Tools (calculator, web search, code interpreter) - Retrieval (BM25, dense, grep, hybrid)

All servers automatically capture telemetry (energy, power, cost, latency).

BaseMCPServer

Bases: ABC

Base class for all MCP servers with automatic telemetry capture.

All subclasses must implement _execute_impl() which performs the actual tool invocation. The base class wraps this with telemetry collection.

Example

class MyTool(BaseMCPServer): def _execute_impl(self, prompt: str, **params) -> MCPToolResult: response = self.api.call(prompt) return MCPToolResult( content=response.text, usage={"prompt_tokens": 100, "completion_tokens": 50}, cost_usd=0.001 )

Source code in intelligence-per-watt/src/ipw/agents/mcp/base.py 
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
class BaseMCPServer(ABC):
    """Base class for all MCP servers with automatic telemetry capture.

    All subclasses must implement _execute_impl() which performs the actual
    tool invocation. The base class wraps this with telemetry collection.

    Example:
        class MyTool(BaseMCPServer):
            def _execute_impl(self, prompt: str, **params) -> MCPToolResult:
                response = self.api.call(prompt)
                return MCPToolResult(
                    content=response.text,
                    usage={"prompt_tokens": 100, "completion_tokens": 50},
                    cost_usd=0.001
                )
    """

    def __init__(
        self,
        name: str,
        telemetry_collector: Optional[Any] = None,
        event_recorder: Optional[Any] = None,
    ):
        """Initialize MCP server.

        Args:
            name: Tool name for logging/tracking
            telemetry_collector: Energy monitor collector. If None, runs without telemetry.
            event_recorder: EventRecorder for per-action tracking. If None, no events recorded.
        """
        self.name = name
        self.telemetry_collector = telemetry_collector
        self.event_recorder = event_recorder

    def execute(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute tool with automatic telemetry capture.

        Args:
            prompt: Input prompt/query for tool
            **params: Additional tool-specific parameters

        Returns:
            MCPToolResult with content, usage, cost, and telemetry samples
        """
        start_time = time.time()

        # Get model info for event recording (subclasses can override these)
        model_id = getattr(self, "model_path", None) or getattr(self, "model_name", self.name)
        model_alias = getattr(self, "model_name", self.name)
        backend = self._get_backend()

        # Record start event
        if self.event_recorder is not None:
            self.event_recorder.record(
                "submodel_call_start",
                model_id=model_id,
                model_alias=model_alias,
                backend=backend,
                tool_name=self.name,
            )

        # Execute with telemetry if available
        if HAS_TELEMETRY and self.telemetry_collector is not None and TelemetrySession is not None:
            with TelemetrySession(self.telemetry_collector) as session:
                result = self._execute_impl(prompt, **params)
                end_time = time.time()
                result.telemetry_samples = list(session.window(start_time, end_time))
                result.latency_seconds = end_time - start_time
        else:
            # Execute without telemetry
            result = self._execute_impl(prompt, **params)
            end_time = time.time()
            result.latency_seconds = end_time - start_time

        # Record end event
        if self.event_recorder is not None:
            self.event_recorder.record(
                "submodel_call_end",
                model_id=model_id,
                model_alias=model_alias,
                backend=backend,
                tool_name=self.name,
                total_tokens=result.usage.get("total_tokens", 0),
                prompt_tokens=result.usage.get("prompt_tokens", 0),
                completion_tokens=result.usage.get("completion_tokens", 0),
                cost_usd=result.cost_usd,
                latency_seconds=result.latency_seconds,
            )

        return result

    def _get_backend(self) -> str:
        """Get the backend type for this server.

        Returns:
            Backend identifier (e.g., 'vllm', 'ollama', 'openai')
        """
        # Default: extract from name if it contains a prefix like "vllm:" or "ollama:"
        if ":" in self.name:
            return self.name.split(":")[0]
        return "unknown"

    @abstractmethod
    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Implement tool execution logic.

        Subclasses must override this to provide actual tool functionality.

        Args:
            prompt: Input prompt/query
            **params: Tool-specific parameters

        Returns:
            MCPToolResult (telemetry_samples and latency_seconds will be added by base class)
        """
        raise NotImplementedError

    def health_check(self) -> bool:
        """Check if tool is available and healthy.

        Returns:
            True if tool is operational, False otherwise
        """
        try:
            result = self.execute("test", timeout=5)
            return result.content is not None
        except Exception:
            return False

    def __repr__(self) -> str:
        return f"{self.__class__.__name__}(name={self.name!r})"

__init__(name, telemetry_collector=None, event_recorder=None)

Initialize MCP server.

Parameters:

Name Type Description Default
name str

Tool name for logging/tracking

 required
telemetry_collector Optional[Any]

Energy monitor collector. If None, runs without telemetry.

 None
event_recorder Optional[Any]

EventRecorder for per-action tracking. If None, no events recorded.

 None
Source code in intelligence-per-watt/src/ipw/agents/mcp/base.py 
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
def __init__(
    self,
    name: str,
    telemetry_collector: Optional[Any] = None,
    event_recorder: Optional[Any] = None,
):
    """Initialize MCP server.

    Args:
        name: Tool name for logging/tracking
        telemetry_collector: Energy monitor collector. If None, runs without telemetry.
        event_recorder: EventRecorder for per-action tracking. If None, no events recorded.
    """
    self.name = name
    self.telemetry_collector = telemetry_collector
    self.event_recorder = event_recorder

execute(prompt, **params)

Execute tool with automatic telemetry capture.

Parameters:

Name Type Description Default
prompt str

Input prompt/query for tool

 required
**params Any

Additional tool-specific parameters

 {}

Returns:

Type Description
MCPToolResult

MCPToolResult with content, usage, cost, and telemetry samples
Source code in intelligence-per-watt/src/ipw/agents/mcp/base.py 
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
def execute(self, prompt: str, **params: Any) -> MCPToolResult:
    """Execute tool with automatic telemetry capture.

    Args:
        prompt: Input prompt/query for tool
        **params: Additional tool-specific parameters

    Returns:
        MCPToolResult with content, usage, cost, and telemetry samples
    """
    start_time = time.time()

    # Get model info for event recording (subclasses can override these)
    model_id = getattr(self, "model_path", None) or getattr(self, "model_name", self.name)
    model_alias = getattr(self, "model_name", self.name)
    backend = self._get_backend()

    # Record start event
    if self.event_recorder is not None:
        self.event_recorder.record(
            "submodel_call_start",
            model_id=model_id,
            model_alias=model_alias,
            backend=backend,
            tool_name=self.name,
        )

    # Execute with telemetry if available
    if HAS_TELEMETRY and self.telemetry_collector is not None and TelemetrySession is not None:
        with TelemetrySession(self.telemetry_collector) as session:
            result = self._execute_impl(prompt, **params)
            end_time = time.time()
            result.telemetry_samples = list(session.window(start_time, end_time))
            result.latency_seconds = end_time - start_time
    else:
        # Execute without telemetry
        result = self._execute_impl(prompt, **params)
        end_time = time.time()
        result.latency_seconds = end_time - start_time

    # Record end event
    if self.event_recorder is not None:
        self.event_recorder.record(
            "submodel_call_end",
            model_id=model_id,
            model_alias=model_alias,
            backend=backend,
            tool_name=self.name,
            total_tokens=result.usage.get("total_tokens", 0),
            prompt_tokens=result.usage.get("prompt_tokens", 0),
            completion_tokens=result.usage.get("completion_tokens", 0),
            cost_usd=result.cost_usd,
            latency_seconds=result.latency_seconds,
        )

    return result

health_check()

Check if tool is available and healthy.

Returns:

Type Description
bool

True if tool is operational, False otherwise
Source code in intelligence-per-watt/src/ipw/agents/mcp/base.py 
185
186
187
188
189
190
191
192
193
194
195
def health_check(self) -> bool:
    """Check if tool is available and healthy.

    Returns:
        True if tool is operational, False otherwise
    """
    try:
        result = self.execute("test", timeout=5)
        return result.content is not None
    except Exception:
        return False

MCPToolResult dataclass

Result from MCP tool execution with telemetry.

Source code in intelligence-per-watt/src/ipw/agents/mcp/base.py 
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
@dataclass
class MCPToolResult:
    """Result from MCP tool execution with telemetry."""

    content: str
    """Response text from tool/model"""

    usage: Dict[str, int] = field(default_factory=dict)
    """Token counts: prompt_tokens, completion_tokens, total_tokens"""

    cost_usd: Optional[float] = None
    """API cost in USD (for cloud APIs)"""

    telemetry_samples: List[Any] = field(default_factory=list)
    """Energy/power/memory readings during execution"""

    latency_seconds: float = 0.0
    """Wall-clock execution time"""

    ttft_seconds: Optional[float] = None
    """Time to first token (for streaming APIs)"""

    metadata: Dict[str, Any] = field(default_factory=dict)
    """Additional tool-specific metadata"""

content instance-attribute

Response text from tool/model

usage = field(default_factory=dict) class-attribute instance-attribute

Token counts: prompt_tokens, completion_tokens, total_tokens

cost_usd = None class-attribute instance-attribute

API cost in USD (for cloud APIs)

telemetry_samples = field(default_factory=list) class-attribute instance-attribute

Energy/power/memory readings during execution

latency_seconds = 0.0 class-attribute instance-attribute

Wall-clock execution time

ttft_seconds = None class-attribute instance-attribute

Time to first token (for streaming APIs)

metadata = field(default_factory=dict) class-attribute instance-attribute

Additional tool-specific metadata

OpenAIMCPServer

Bases: BaseMCPServer

MCP server for OpenAI models with automatic cost tracking.

Tracks API costs based on token usage and current pricing.

Example

server = OpenAIMCPServer( model_name="gpt-4o", api_key=os.getenv("OPENAI_API_KEY") )

result = server.execute("Explain quantum computing") print(result.content) print(f"Cost: ${result.cost_usd:.4f}")

Source code in intelligence-per-watt/src/ipw/agents/mcp/openai_server.py 
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
class OpenAIMCPServer(BaseMCPServer):
    """MCP server for OpenAI models with automatic cost tracking.

    Tracks API costs based on token usage and current pricing.

    Example:
        server = OpenAIMCPServer(
            model_name="gpt-4o",
            api_key=os.getenv("OPENAI_API_KEY")
        )

        result = server.execute("Explain quantum computing")
        print(result.content)
        print(f"Cost: ${result.cost_usd:.4f}")
    """

    def __init__(
        self,
        model_name: str,
        api_key: Optional[str] = None,
        telemetry_collector: Optional[Any] = None,
        event_recorder: Optional[Any] = None,
        **openai_params: Any,
    ):
        """Initialize OpenAI MCP server.

        Args:
            model_name: OpenAI model name (e.g., "gpt-4o", "gpt-5-mini-2025-08-07")
            api_key: OpenAI API key (or set OPENAI_API_KEY env var)
            telemetry_collector: Energy monitor collector
            event_recorder: EventRecorder for per-action tracking
            **openai_params: Additional OpenAI parameters (temperature, max_tokens, etc.)
        """
        super().__init__(
            name=f"openai:{model_name}",
            telemetry_collector=telemetry_collector,
            event_recorder=event_recorder,
        )

        self.model_name = model_name
        self.openai_params = openai_params

        # Lazy import: openai is optional
        try:
            from openai import OpenAI
        except ImportError:
            raise ImportError(
                "openai package is required for OpenAIMCPServer. "
                "Install with: pip install openai"
            )

        # Initialize OpenAI client
        if api_key is None:
            api_key = os.getenv("OPENAI_API_KEY")
        self._client = OpenAI(api_key=api_key)

        # Get pricing for this model
        self.pricing = OPENAI_PRICING.get(model_name)
        if not self.pricing:
            print(f"Warning: No pricing info for {model_name}, using gpt-4o rates")
            self.pricing = OPENAI_PRICING["gpt-4o"]

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute OpenAI API call with cost tracking."""
        from openai import OpenAIError

        # Merge default params with request params
        payload = {**self.openai_params, **params}
        payload["model"] = self.model_name
        payload["messages"] = [{"role": "user", "content": prompt}]
        payload["stream"] = True

        # GPT-5+ models use max_completion_tokens instead of max_tokens
        if "max_tokens" in payload and self.model_name.startswith("gpt-5"):
            payload["max_completion_tokens"] = payload.pop("max_tokens")

        # Call OpenAI API
        start = time.perf_counter()
        try:
            stream = self._client.chat.completions.create(**payload)
        except OpenAIError as exc:
            raise RuntimeError(f"OpenAI error for {self.model_name}: {exc}") from exc

        # Consume stream and collect response
        content_chunks: list[str] = []
        ttft_ms: Optional[float] = None
        usage = None

        for chunk in stream:
            if chunk.choices and len(chunk.choices) > 0:
                delta = chunk.choices[0].delta
                if delta.content:
                    if ttft_ms is None:
                        ttft_ms = (time.perf_counter() - start) * 1000
                    content_chunks.append(delta.content)

            # Last chunk contains usage
            if hasattr(chunk, "usage") and chunk.usage:
                usage = chunk.usage

        content = "".join(content_chunks)

        # Extract token counts
        if usage:
            prompt_tokens = usage.prompt_tokens
            completion_tokens = usage.completion_tokens
            total_tokens = usage.total_tokens
        else:
            prompt_tokens = len(prompt.split())
            completion_tokens = len(content.split())
            total_tokens = prompt_tokens + completion_tokens

        # Calculate cost
        cost_usd = calculate_cost("openai", self.model_name, prompt_tokens, completion_tokens)

        return MCPToolResult(
            content=content,
            usage={
                "prompt_tokens": prompt_tokens,
                "completion_tokens": completion_tokens,
                "total_tokens": total_tokens,
            },
            cost_usd=cost_usd,
            ttft_seconds=(ttft_ms / 1000.0) if ttft_ms else None,
            metadata={
                "model": self.model_name,
                "backend": "openai",
                "pricing_input_per_1m": self.pricing["input"],
                "pricing_output_per_1m": self.pricing["output"],
            },
        )

    def health_check(self) -> bool:
        """Check if OpenAI API is accessible."""
        try:
            response = self._client.chat.completions.create(
                model=self.model_name,
                messages=[{"role": "user", "content": "test"}],
                max_tokens=1,
            )
            return response is not None
        except Exception:
            return False

    def list_available_models(self) -> list[str]:
        """List all available OpenAI models."""
        from openai import OpenAIError
        try:
            models = self._client.models.list()
            return [model.id for model in models.data if model.id.startswith("gpt")]
        except OpenAIError as exc:
            raise RuntimeError(f"Failed to list OpenAI models: {exc}") from exc

__init__(model_name, api_key=None, telemetry_collector=None, event_recorder=None, **openai_params)

Initialize OpenAI MCP server.

Parameters:

Name Type Description Default
model_name str

OpenAI model name (e.g., "gpt-4o", "gpt-5-mini-2025-08-07")

 required
api_key Optional[str]

OpenAI API key (or set OPENAI_API_KEY env var)

 None
telemetry_collector Optional[Any]

Energy monitor collector

 None
event_recorder Optional[Any]

EventRecorder for per-action tracking

 None
**openai_params Any

Additional OpenAI parameters (temperature, max_tokens, etc.)

 {}
Source code in intelligence-per-watt/src/ipw/agents/mcp/openai_server.py 
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
def __init__(
    self,
    model_name: str,
    api_key: Optional[str] = None,
    telemetry_collector: Optional[Any] = None,
    event_recorder: Optional[Any] = None,
    **openai_params: Any,
):
    """Initialize OpenAI MCP server.

    Args:
        model_name: OpenAI model name (e.g., "gpt-4o", "gpt-5-mini-2025-08-07")
        api_key: OpenAI API key (or set OPENAI_API_KEY env var)
        telemetry_collector: Energy monitor collector
        event_recorder: EventRecorder for per-action tracking
        **openai_params: Additional OpenAI parameters (temperature, max_tokens, etc.)
    """
    super().__init__(
        name=f"openai:{model_name}",
        telemetry_collector=telemetry_collector,
        event_recorder=event_recorder,
    )

    self.model_name = model_name
    self.openai_params = openai_params

    # Lazy import: openai is optional
    try:
        from openai import OpenAI
    except ImportError:
        raise ImportError(
            "openai package is required for OpenAIMCPServer. "
            "Install with: pip install openai"
        )

    # Initialize OpenAI client
    if api_key is None:
        api_key = os.getenv("OPENAI_API_KEY")
    self._client = OpenAI(api_key=api_key)

    # Get pricing for this model
    self.pricing = OPENAI_PRICING.get(model_name)
    if not self.pricing:
        print(f"Warning: No pricing info for {model_name}, using gpt-4o rates")
        self.pricing = OPENAI_PRICING["gpt-4o"]

health_check()

Check if OpenAI API is accessible.

Source code in intelligence-per-watt/src/ipw/agents/mcp/openai_server.py 
145
146
147
148
149
150
151
152
153
154
155
def health_check(self) -> bool:
    """Check if OpenAI API is accessible."""
    try:
        response = self._client.chat.completions.create(
            model=self.model_name,
            messages=[{"role": "user", "content": "test"}],
            max_tokens=1,
        )
        return response is not None
    except Exception:
        return False

list_available_models()

List all available OpenAI models.

Source code in intelligence-per-watt/src/ipw/agents/mcp/openai_server.py 
157
158
159
160
161
162
163
164
def list_available_models(self) -> list[str]:
    """List all available OpenAI models."""
    from openai import OpenAIError
    try:
        models = self._client.models.list()
        return [model.id for model in models.data if model.id.startswith("gpt")]
    except OpenAIError as exc:
        raise RuntimeError(f"Failed to list OpenAI models: {exc}") from exc

AnthropicMCPServer

Bases: BaseMCPServer

MCP server for Anthropic Claude models with cost tracking.

Example

server = AnthropicMCPServer( model_name="claude-sonnet-4-5-20250929", api_key=os.getenv("ANTHROPIC_API_KEY") )

result = server.execute("Write a haiku about AI") print(result.content) print(f"Cost: ${result.cost_usd:.4f}")

Source code in intelligence-per-watt/src/ipw/agents/mcp/anthropic_server.py 
 12
 13
 14
 15
 16
 17
 18
 19
 20
 21
 22
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
class AnthropicMCPServer(BaseMCPServer):
    """MCP server for Anthropic Claude models with cost tracking.

    Example:
        server = AnthropicMCPServer(
            model_name="claude-sonnet-4-5-20250929",
            api_key=os.getenv("ANTHROPIC_API_KEY")
        )

        result = server.execute("Write a haiku about AI")
        print(result.content)
        print(f"Cost: ${result.cost_usd:.4f}")
    """

    def __init__(
        self,
        model_name: str,
        api_key: Optional[str] = None,
        telemetry_collector: Optional[Any] = None,
        **anthropic_params: Any,
    ):
        """Initialize Anthropic MCP server.

        Args:
            model_name: Claude model name
            api_key: Anthropic API key (or set ANTHROPIC_API_KEY env var)
            telemetry_collector: Energy monitor collector
            **anthropic_params: Additional params (temperature, max_tokens, etc.)
        """
        super().__init__(
            name=f"anthropic:{model_name}",
            telemetry_collector=telemetry_collector,
        )

        self.model_name = model_name
        self.anthropic_params = anthropic_params

        # Lazy import: anthropic is optional
        try:
            from anthropic import Anthropic
        except ImportError:
            raise ImportError(
                "anthropic package is required for AnthropicMCPServer. "
                "Install with: pip install anthropic"
            )

        # Initialize Anthropic client
        self._client = Anthropic(api_key=api_key)

        # Get pricing for this model
        self.pricing = ANTHROPIC_PRICING.get(model_name)
        if not self.pricing:
            print(f"Warning: No pricing info for {model_name}, using Sonnet 4.5 rates")
            self.pricing = ANTHROPIC_PRICING["claude-sonnet-4-5-20250929"]

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute Anthropic API call with cost tracking."""
        from anthropic import AnthropicError

        # Merge default params with request params
        payload = {**self.anthropic_params, **params}
        payload["model"] = self.model_name
        payload["messages"] = [{"role": "user", "content": prompt}]
        payload["max_tokens"] = payload.get("max_tokens", 4096)
        payload["stream"] = True

        # Call Anthropic API
        start = time.perf_counter()
        try:
            stream = self._client.messages.create(**payload)
        except AnthropicError as exc:
            raise RuntimeError(f"Anthropic error for {self.model_name}: {exc}") from exc

        # Consume stream and collect response
        content_chunks: list[str] = []
        ttft_ms: Optional[float] = None
        input_tokens = 0
        output_tokens = 0

        with stream as event_stream:
            for event in event_stream:
                if event.type == "content_block_delta":
                    if hasattr(event.delta, "text"):
                        if ttft_ms is None:
                            ttft_ms = (time.perf_counter() - start) * 1000
                        content_chunks.append(event.delta.text)

                elif event.type == "message_start":
                    if hasattr(event.message, "usage"):
                        input_tokens = event.message.usage.input_tokens

                elif event.type == "message_delta":
                    if hasattr(event, "usage"):
                        output_tokens = event.usage.output_tokens

        content = "".join(content_chunks)

        # Calculate cost
        cost_usd = calculate_cost("anthropic", self.model_name, input_tokens, output_tokens)

        return MCPToolResult(
            content=content,
            usage={
                "prompt_tokens": input_tokens,
                "completion_tokens": output_tokens,
                "total_tokens": input_tokens + output_tokens,
            },
            cost_usd=cost_usd,
            ttft_seconds=(ttft_ms / 1000.0) if ttft_ms else None,
            metadata={
                "model": self.model_name,
                "backend": "anthropic",
                "pricing_input_per_1m": self.pricing["input"],
                "pricing_output_per_1m": self.pricing["output"],
            },
        )

    def health_check(self) -> bool:
        """Check if Anthropic API is accessible."""
        try:
            response = self._client.messages.create(
                model=self.model_name,
                messages=[{"role": "user", "content": "test"}],
                max_tokens=1,
            )
            return response is not None
        except Exception:
            return False

__init__(model_name, api_key=None, telemetry_collector=None, **anthropic_params)

Initialize Anthropic MCP server.

Parameters:

Name Type Description Default
model_name str

Claude model name

 required
api_key Optional[str]

Anthropic API key (or set ANTHROPIC_API_KEY env var)

 None
telemetry_collector Optional[Any]

Energy monitor collector

 None
**anthropic_params Any

Additional params (temperature, max_tokens, etc.)

 {}
Source code in intelligence-per-watt/src/ipw/agents/mcp/anthropic_server.py 
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
def __init__(
    self,
    model_name: str,
    api_key: Optional[str] = None,
    telemetry_collector: Optional[Any] = None,
    **anthropic_params: Any,
):
    """Initialize Anthropic MCP server.

    Args:
        model_name: Claude model name
        api_key: Anthropic API key (or set ANTHROPIC_API_KEY env var)
        telemetry_collector: Energy monitor collector
        **anthropic_params: Additional params (temperature, max_tokens, etc.)
    """
    super().__init__(
        name=f"anthropic:{model_name}",
        telemetry_collector=telemetry_collector,
    )

    self.model_name = model_name
    self.anthropic_params = anthropic_params

    # Lazy import: anthropic is optional
    try:
        from anthropic import Anthropic
    except ImportError:
        raise ImportError(
            "anthropic package is required for AnthropicMCPServer. "
            "Install with: pip install anthropic"
        )

    # Initialize Anthropic client
    self._client = Anthropic(api_key=api_key)

    # Get pricing for this model
    self.pricing = ANTHROPIC_PRICING.get(model_name)
    if not self.pricing:
        print(f"Warning: No pricing info for {model_name}, using Sonnet 4.5 rates")
        self.pricing = ANTHROPIC_PRICING["claude-sonnet-4-5-20250929"]

health_check()

Check if Anthropic API is accessible.

Source code in intelligence-per-watt/src/ipw/agents/mcp/anthropic_server.py 
129
130
131
132
133
134
135
136
137
138
139
def health_check(self) -> bool:
    """Check if Anthropic API is accessible."""
    try:
        response = self._client.messages.create(
            model=self.model_name,
            messages=[{"role": "user", "content": "test"}],
            max_tokens=1,
        )
        return response is not None
    except Exception:
        return False

OpenRouterMCPServer

Bases: BaseMCPServer

MCP server for OpenRouter with automatic cost tracking.

OpenRouter provides a unified API for accessing many LLM providers. Uses OpenAI-compatible API format.

Example

server = OpenRouterMCPServer( model_name="google/gemini-2.5-flash", api_key=os.getenv("OPENROUTER_API_KEY") )

result = server.execute("Explain quantum computing") print(result.content) print(f"Cost: ${result.cost_usd:.6f}")

Source code in intelligence-per-watt/src/ipw/agents/mcp/openrouter_server.py 
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
class OpenRouterMCPServer(BaseMCPServer):
    """MCP server for OpenRouter with automatic cost tracking.

    OpenRouter provides a unified API for accessing many LLM providers.
    Uses OpenAI-compatible API format.

    Example:
        server = OpenRouterMCPServer(
            model_name="google/gemini-2.5-flash",
            api_key=os.getenv("OPENROUTER_API_KEY")
        )

        result = server.execute("Explain quantum computing")
        print(result.content)
        print(f"Cost: ${result.cost_usd:.6f}")
    """

    # Approximate pricing per 1M tokens (varies by model)
    # Source: https://openrouter.ai/models
    PRICING = {
        # Meta Llama
        "meta-llama/llama-3.3-70b-instruct": {"input": 0.40, "output": 0.40},
        "meta-llama/llama-3.1-405b-instruct": {"input": 2.00, "output": 2.00},
        "meta-llama/llama-3.1-70b-instruct": {"input": 0.40, "output": 0.40},
        "meta-llama/llama-3.1-8b-instruct": {"input": 0.05, "output": 0.05},
        # Qwen - General
        "qwen/qwen-2.5-72b-instruct": {"input": 0.35, "output": 0.40},
        "qwen/qwen-2.5-32b-instruct": {"input": 0.15, "output": 0.15},
        "qwen/qwq-32b": {"input": 0.15, "output": 0.15},
        "qwen/qwen-2.5-coder-32b-instruct": {"input": 0.15, "output": 0.15},
        # Qwen3
        "qwen/qwen3-32b": {"input": 0.08, "output": 0.24},
        "qwen/qwen3-coder-next": {"input": 0.20, "output": 1.50},
        "qwen/qwen3-coder-plus": {"input": 1.00, "output": 5.00},
        "qwen/qwen3-max": {"input": 1.20, "output": 6.00},
        "qwen/qwen3-next-80b-a3b-instruct": {"input": 0.09, "output": 1.10},
        # Math specialist
        "z-ai/glm-4.7": {"input": 0.40, "output": 1.50},
        "qwen/qwen2.5-math-72b-instruct": {"input": 0.40, "output": 0.40},
        "qwen/qwen2.5-coder-32b-instruct": {"input": 0.15, "output": 0.15},
        # DeepSeek
        "deepseek/deepseek-r1": {"input": 0.70, "output": 2.50},
        "deepseek/deepseek-r1-0528": {"input": 0.40, "output": 1.75},
        "deepseek/deepseek-v3.2": {"input": 0.25, "output": 0.38},
        "deepseek/deepseek-v3.1-terminus": {"input": 0.21, "output": 0.79},
        "deepseek/deepseek-chat-v3-0324": {"input": 0.14, "output": 0.28},
        "deepseek/deepseek-chat": {"input": 0.14, "output": 0.28},
        "deepseek/deepseek-coder-v2": {"input": 0.14, "output": 0.28},
        # Mistral
        "mistralai/mistral-large-2411": {"input": 2.00, "output": 6.00},
        "mistralai/mistral-small-2501": {"input": 0.10, "output": 0.30},
        "mistralai/codestral-2501": {"input": 0.30, "output": 0.90},
        # Google Gemini (via OpenRouter)
        "google/gemini-2.5-flash": {"input": 0.30, "output": 2.50},
        "google/gemini-2.5-pro": {"input": 1.25, "output": 5.00},
        "google/gemini-3-flash-preview": {"input": 0.50, "output": 3.00},
        "z-ai/glm-4.7-flash": {"input": 0.07, "output": 0.40},
        # Small/cheap models
        "openai/gpt-oss-20b": {"input": 0.02, "output": 0.10},
        "deepseek/deepseek-r1-distill-qwen-1.5b": {"input": 0.02, "output": 0.05},
        "google/gemma-3-4b-it": {"input": 0.02, "output": 0.07},
    }

    # Default fallback pricing
    DEFAULT_PRICING = {"input": 1.00, "output": 3.00}

    def __init__(
        self,
        model_name: str,
        api_key: Optional[str] = None,
        telemetry_collector: Optional[Any] = None,
        site_url: Optional[str] = None,
        app_name: Optional[str] = None,
        **openai_params: Any,
    ):
        """Initialize OpenRouter MCP server.

        Args:
            model_name: Model identifier (e.g., "google/gemini-2.5-flash")
            api_key: OpenRouter API key (or set OPENROUTER_API_KEY env var)
            telemetry_collector: Energy monitor collector
            site_url: Your site URL for OpenRouter rankings
            app_name: Your app name for OpenRouter rankings
            **openai_params: Additional parameters (temperature, max_tokens, etc.)
        """
        super().__init__(
            name=f"openrouter:{model_name}",
            telemetry_collector=telemetry_collector,
        )

        self.model_name = model_name
        self.openai_params = openai_params

        # Get API key
        api_key = api_key or os.environ.get("OPENROUTER_API_KEY")
        if not api_key:
            raise ValueError(
                "OpenRouter API key required. Set OPENROUTER_API_KEY env var "
                "or pass api_key parameter."
            )

        # Lazy import: openai is optional
        try:
            from openai import OpenAI
        except ImportError:
            raise ImportError(
                "openai package is required for OpenRouterMCPServer. "
                "Install with: pip install openai"
            )

        # Initialize OpenAI-compatible client with OpenRouter base URL
        self._client = OpenAI(
            base_url="https://openrouter.ai/api/v1",
            api_key=api_key,
            default_headers={
                "HTTP-Referer": site_url or "https://github.com/ipw",
                "X-Title": app_name or "IPW Orchestrator",
            },
        )

        # Get pricing for this model
        self.pricing = self.PRICING.get(model_name, self.DEFAULT_PRICING)
        if model_name not in self.PRICING:
            print(f"Warning: No pricing info for {model_name}, using default rates")

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute OpenRouter API call with cost tracking."""
        from openai import OpenAIError

        # Merge default params with request params
        payload = {**self.openai_params, **params}
        payload["model"] = self.model_name
        payload["messages"] = [{"role": "user", "content": prompt}]
        payload["stream"] = True

        # Call OpenRouter API
        start = time.perf_counter()
        try:
            stream = self._client.chat.completions.create(**payload)
        except OpenAIError as exc:
            raise RuntimeError(f"OpenRouter error for {self.model_name}: {exc}") from exc

        # Consume stream and collect response
        content_chunks: list[str] = []
        ttft_ms: Optional[float] = None
        usage = None

        for chunk in stream:
            if chunk.choices and len(chunk.choices) > 0:
                delta = chunk.choices[0].delta
                if delta.content:
                    if ttft_ms is None:
                        ttft_ms = (time.perf_counter() - start) * 1000
                    content_chunks.append(delta.content)

            # Last chunk may contain usage
            if hasattr(chunk, "usage") and chunk.usage:
                usage = chunk.usage

        content = "".join(content_chunks)

        # Extract token counts
        if usage:
            prompt_tokens = usage.prompt_tokens
            completion_tokens = usage.completion_tokens
            total_tokens = usage.total_tokens
        else:
            prompt_tokens = int(len(prompt.split()) * 1.3)
            completion_tokens = int(len(content.split()) * 1.3)
            total_tokens = prompt_tokens + completion_tokens

        # Calculate cost based on token usage
        cost_usd = self._calculate_cost(prompt_tokens, completion_tokens)

        return MCPToolResult(
            content=content,
            usage={
                "prompt_tokens": prompt_tokens,
                "completion_tokens": completion_tokens,
                "total_tokens": total_tokens,
            },
            cost_usd=cost_usd,
            ttft_seconds=(ttft_ms / 1000.0) if ttft_ms else None,
            metadata={
                "model": self.model_name,
                "backend": "openrouter",
                "pricing_input_per_1m": self.pricing["input"],
                "pricing_output_per_1m": self.pricing["output"],
            },
        )

    def _calculate_cost(self, prompt_tokens: int, completion_tokens: int) -> float:
        """Calculate API cost in USD."""
        input_cost = (prompt_tokens / 1_000_000) * self.pricing["input"]
        output_cost = (completion_tokens / 1_000_000) * self.pricing["output"]
        return input_cost + output_cost

    def health_check(self) -> bool:
        """Check if OpenRouter API is accessible."""
        try:
            response = self._client.chat.completions.create(
                model=self.model_name,
                messages=[{"role": "user", "content": "test"}],
                max_tokens=1,
            )
            return response is not None
        except Exception:
            return False

    @classmethod
    def list_popular_models(cls) -> list[str]:
        """List popular models available on OpenRouter."""
        return list(cls.PRICING.keys())

__init__(model_name, api_key=None, telemetry_collector=None, site_url=None, app_name=None, **openai_params)

Initialize OpenRouter MCP server.

Parameters:

Name Type Description Default
model_name str

Model identifier (e.g., "google/gemini-2.5-flash")

 required
api_key Optional[str]

OpenRouter API key (or set OPENROUTER_API_KEY env var)

 None
telemetry_collector Optional[Any]

Energy monitor collector

 None
site_url Optional[str]

Your site URL for OpenRouter rankings

 None
app_name Optional[str]

Your app name for OpenRouter rankings

 None
**openai_params Any

Additional parameters (temperature, max_tokens, etc.)

 {}
Source code in intelligence-per-watt/src/ipw/agents/mcp/openrouter_server.py 
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
def __init__(
    self,
    model_name: str,
    api_key: Optional[str] = None,
    telemetry_collector: Optional[Any] = None,
    site_url: Optional[str] = None,
    app_name: Optional[str] = None,
    **openai_params: Any,
):
    """Initialize OpenRouter MCP server.

    Args:
        model_name: Model identifier (e.g., "google/gemini-2.5-flash")
        api_key: OpenRouter API key (or set OPENROUTER_API_KEY env var)
        telemetry_collector: Energy monitor collector
        site_url: Your site URL for OpenRouter rankings
        app_name: Your app name for OpenRouter rankings
        **openai_params: Additional parameters (temperature, max_tokens, etc.)
    """
    super().__init__(
        name=f"openrouter:{model_name}",
        telemetry_collector=telemetry_collector,
    )

    self.model_name = model_name
    self.openai_params = openai_params

    # Get API key
    api_key = api_key or os.environ.get("OPENROUTER_API_KEY")
    if not api_key:
        raise ValueError(
            "OpenRouter API key required. Set OPENROUTER_API_KEY env var "
            "or pass api_key parameter."
        )

    # Lazy import: openai is optional
    try:
        from openai import OpenAI
    except ImportError:
        raise ImportError(
            "openai package is required for OpenRouterMCPServer. "
            "Install with: pip install openai"
        )

    # Initialize OpenAI-compatible client with OpenRouter base URL
    self._client = OpenAI(
        base_url="https://openrouter.ai/api/v1",
        api_key=api_key,
        default_headers={
            "HTTP-Referer": site_url or "https://github.com/ipw",
            "X-Title": app_name or "IPW Orchestrator",
        },
    )

    # Get pricing for this model
    self.pricing = self.PRICING.get(model_name, self.DEFAULT_PRICING)
    if model_name not in self.PRICING:
        print(f"Warning: No pricing info for {model_name}, using default rates")

health_check()

Check if OpenRouter API is accessible.

Source code in intelligence-per-watt/src/ipw/agents/mcp/openrouter_server.py 
220
221
222
223
224
225
226
227
228
229
230
def health_check(self) -> bool:
    """Check if OpenRouter API is accessible."""
    try:
        response = self._client.chat.completions.create(
            model=self.model_name,
            messages=[{"role": "user", "content": "test"}],
            max_tokens=1,
        )
        return response is not None
    except Exception:
        return False

List popular models available on OpenRouter.

Source code in intelligence-per-watt/src/ipw/agents/mcp/openrouter_server.py 
232
233
234
235
@classmethod
def list_popular_models(cls) -> list[str]:
    """List popular models available on OpenRouter."""
    return list(cls.PRICING.keys())

VLLMMCPServer

Bases: BaseMCPServer

MCP server for vLLM-served models.

vLLM provides an OpenAI-compatible API for serving large open-source models with optimizations like PagedAttention, continuous batching, and tensor parallelism.

Supported model categories: - General: Qwen3-32B, Qwen3-8B, Llama-3.3-70B-Instruct - Math specialist: Qwen2.5-Math-72B, Qwen2.5-Math-7B - Code specialist: Qwen2.5-Coder-32B, DeepSeek-Coder-V2

Example

Start vLLM server externally:

vllm serve Qwen/Qwen3-32B --tensor-parallel-size 4 --port 8000

server = VLLMMCPServer(model_name="qwen3-32b") result = server.execute("Explain quantum computing")

Source code in intelligence-per-watt/src/ipw/agents/mcp/vllm_server.py 
 23
 24
 25
 26
 27
 28
 29
 30
 31
 32
 33
 34
 35
 36
 37
 38
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
class VLLMMCPServer(BaseMCPServer):
    """MCP server for vLLM-served models.

    vLLM provides an OpenAI-compatible API for serving large open-source
    models with optimizations like PagedAttention, continuous batching,
    and tensor parallelism.

    Supported model categories:
    - General: Qwen3-32B, Qwen3-8B, Llama-3.3-70B-Instruct
    - Math specialist: Qwen2.5-Math-72B, Qwen2.5-Math-7B
    - Code specialist: Qwen2.5-Coder-32B, DeepSeek-Coder-V2

    Example:
        # Start vLLM server externally:
        # vllm serve Qwen/Qwen3-32B --tensor-parallel-size 4 --port 8000

        server = VLLMMCPServer(model_name="qwen3-32b")
        result = server.execute("Explain quantum computing")
    """

    # Model name aliases to full HuggingFace paths
    SUPPORTED_MODELS: Dict[str, str] = {
        # General purpose
        "qwen3-32b": "Qwen/Qwen3-32B",
        "qwen3-8b": "Qwen/Qwen3-8B",
        "llama-70b": "meta-llama/Llama-3.3-70B-Instruct",
        "llama-8b": "meta-llama/Llama-3.1-8B-Instruct",
        # Math specialists
        "glm-4.7": "THUDM/glm-4-9b-chat",
        "qwen-math-7b": "Qwen/Qwen2.5-Math-7B-Instruct",
        "qwen-math-1.5b": "Qwen/Qwen2.5-Math-1.5B-Instruct",
        # Code specialists
        "qwen3-coder-plus": "Qwen/Qwen3-Coder-Plus",
        "qwen-coder-7b": "Qwen/Qwen2.5-Coder-7B-Instruct",
        # MoE models
        "glm-4.7-flash": "zai-org/GLM-4.7-Flash",
    }

    # Estimated costs per 1M tokens (local compute, GPU rental approximation)
    MODEL_COSTS: Dict[str, Dict[str, float]] = {
        "qwen3-32b": {"prompt": 0.50, "completion": 0.50},
        "qwen3-8b": {"prompt": 0.10, "completion": 0.10},
        "llama-70b": {"prompt": 1.00, "completion": 1.00},
        "llama-8b": {"prompt": 0.10, "completion": 0.10},
        "glm-4.7": {"prompt": 1.00, "completion": 1.00},
        "qwen-math-7b": {"prompt": 0.10, "completion": 0.10},
        "qwen-math-1.5b": {"prompt": 0.02, "completion": 0.02},
        "qwen3-coder-plus": {"prompt": 0.50, "completion": 0.50},
        "qwen-coder-7b": {"prompt": 0.10, "completion": 0.10},
        "glm-4.7-flash": {"prompt": 0.30, "completion": 0.30},
    }

    def __init__(
        self,
        model_name: str,
        vllm_url: str = "http://localhost:8000",
        api_key: Optional[str] = None,
        telemetry_collector: Optional[Any] = None,
        event_recorder: Optional[Any] = None,
        **vllm_params: Any,
    ):
        """Initialize vLLM server connection.

        Args:
            model_name: Model alias (e.g., 'qwen3-32b') or full HF path
            vllm_url: URL of the vLLM server (default: localhost:8000)
            api_key: Optional API key for authenticated endpoints
            telemetry_collector: Energy monitor collector
            event_recorder: EventRecorder for per-action tracking
            **vllm_params: Default parameters (max_tokens, temperature, top_p, etc.)
        """
        super().__init__(
            name=f"vllm:{model_name}",
            telemetry_collector=telemetry_collector,
            event_recorder=event_recorder,
        )

        self.model_name = model_name
        self.model_path = self.SUPPORTED_MODELS.get(model_name, model_name)
        self.vllm_url = vllm_url.rstrip("/")
        self.api_key = api_key or os.environ.get("VLLM_API_KEY")
        self.vllm_params = vllm_params

        # Cost estimation (per 1M tokens)
        self.cost_per_1m = self.MODEL_COSTS.get(
            model_name,
            {"prompt": 0.0, "completion": 0.0}
        )

        # Query server's actual max_model_len to handle validation
        self._server_max_model_len: Optional[int] = None

    def _get_server_max_model_len(self) -> Optional[int]:
        """Query server's actual max_model_len from /v1/models endpoint."""
        if self._server_max_model_len is not None:
            return self._server_max_model_len

        try:
            with httpx.Client(timeout=5.0) as client:
                response = client.get(f"{self.vllm_url}/v1/models")
                if response.status_code == 200:
                    models = response.json().get("data", [])
                    for model in models:
                        if model.get("id") == self.model_path or model.get("id") == self.model_name:
                            self._server_max_model_len = model.get("max_model_len")
                            return self._server_max_model_len
        except Exception:
            pass
        return None

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute inference via vLLM's OpenAI-compatible API."""
        global _retry_warn_count

        # Merge default params with per-request params
        merged_params = {**self.vllm_params, **params}

        max_tokens = merged_params.get("max_tokens", 8192)
        temperature = merged_params.get("temperature", 0.7)
        top_p = merged_params.get("top_p", 0.9)
        system_prompt = merged_params.get("system_prompt")

        # Build messages
        messages = []
        if system_prompt:
            messages.append({"role": "system", "content": system_prompt})
        messages.append({"role": "user", "content": prompt})

        # Build request
        headers = {"Content-Type": "application/json"}
        if self.api_key:
            headers["Authorization"] = f"Bearer {self.api_key}"

        original_max_tokens = max_tokens

        payload = {
            "model": self.model_path,
            "messages": messages,
            "max_tokens": max_tokens,
            "temperature": temperature,
            "top_p": top_p,
        }

        try:
            with httpx.Client(timeout=240.0) as client:
                response = client.post(
                    f"{self.vllm_url}/v1/chat/completions",
                    headers=headers,
                    json=payload,
                )
                response.raise_for_status()
                data = response.json()

            # Extract response
            content = data["choices"][0]["message"]["content"]
            usage = data.get("usage", {})

            prompt_tokens = usage.get("prompt_tokens", 0)
            completion_tokens = usage.get("completion_tokens", 0)

            # Calculate cost
            cost_usd = (
                (prompt_tokens / 1_000_000) * self.cost_per_1m["prompt"] +
                (completion_tokens / 1_000_000) * self.cost_per_1m["completion"]
            )

            return MCPToolResult(
                content=content,
                usage={
                    "prompt_tokens": prompt_tokens,
                    "completion_tokens": completion_tokens,
                    "total_tokens": prompt_tokens + completion_tokens,
                },
                cost_usd=cost_usd,
                metadata={
                    "model": self.model_path,
                    "server": "vllm",
                    "finish_reason": data["choices"][0].get("finish_reason"),
                },
            )

        except httpx.ConnectError:
            return MCPToolResult(
                content=f"Error: Cannot connect to vLLM server at {self.vllm_url}. "
                        f"Please ensure vLLM is running with the model {self.model_path}.",
                usage={},
                cost_usd=0.0,
                metadata={"error": "connection_failed"},
            )
        except httpx.ReadTimeout:
            return MCPToolResult(
                content=f"Error: vLLM request timed out after 240s. "
                        f"The model may be overloaded or the response is too long.",
                usage={},
                cost_usd=0.0,
                metadata={"error": "timeout"},
            )
        except httpx.HTTPStatusError as e:
            # Handle max_tokens validation error - try with capped value
            error_text = e.response.text
            error_text_lower = error_text.lower()

            if e.response.status_code == 400 and ("max_tokens" in error_text_lower or "max_completion_tokens" in error_text_lower) and "too large" in error_text_lower:
                match = re.search(r"maximum context length is (\d+)", error_text_lower)
                input_match = re.search(r"your request has (\d+) input", error_text_lower)

                if match:
                    validation_limit = int(match.group(1))
                    actual_input_tokens = int(input_match.group(1)) if input_match else None

                    if actual_input_tokens:
                        capped_max_tokens = max(1, int(validation_limit - actual_input_tokens - 100))
                    else:
                        estimated_prompt_tokens = len(prompt.split()) * 1.3
                        capped_max_tokens = max(1, int(validation_limit - estimated_prompt_tokens - 100))

                    if capped_max_tokens < max_tokens and capped_max_tokens > 0:
                        _retry_warn_count += 1
                        if _retry_warn_count == 1 or _retry_warn_count % 100 == 0:
                            warnings.warn(
                                f"vLLM validation rejected max_tokens={max_tokens} "
                                f"(validation uses model config limit: {validation_limit}). "
                                f"Retrying with max_tokens={capped_max_tokens}. "
                                f"(Total retries: {_retry_warn_count})",
                                UserWarning,
                                stacklevel=2,
                            )

                        payload["max_tokens"] = capped_max_tokens
                        try:
                            with httpx.Client(timeout=240.0) as client:
                                retry_response = client.post(
                                    f"{self.vllm_url}/v1/chat/completions",
                                    headers=headers,
                                    json=payload,
                                )
                                retry_response.raise_for_status()
                                retry_data = retry_response.json()

                            content = retry_data["choices"][0]["message"]["content"]
                            usage = retry_data.get("usage", {})
                            prompt_tokens = usage.get("prompt_tokens", 0)
                            completion_tokens = usage.get("completion_tokens", 0)
                            cost_usd = (
                                (prompt_tokens / 1_000_000) * self.cost_per_1m["prompt"] +
                                (completion_tokens / 1_000_000) * self.cost_per_1m["completion"]
                            )
                            return MCPToolResult(
                                content=content,
                                usage={
                                    "prompt_tokens": prompt_tokens,
                                    "completion_tokens": completion_tokens,
                                    "total_tokens": prompt_tokens + completion_tokens,
                                },
                                cost_usd=cost_usd,
                                metadata={
                                    "model": self.model_path,
                                    "server": "vllm",
                                    "finish_reason": retry_data["choices"][0].get("finish_reason"),
                                    "max_tokens_capped": True,
                                    "original_max_tokens": original_max_tokens,
                                },
                            )
                        except httpx.HTTPStatusError as retry_e:
                            return MCPToolResult(
                                content=f"Error: vLLM server returned {retry_e.response.status_code}: {retry_e.response.text}",
                                usage={},
                                cost_usd=0.0,
                                metadata={"error": f"http_{retry_e.response.status_code}", "retry_failed": True},
                            )
                        except Exception as retry_e:
                            return MCPToolResult(
                                content=f"Error during retry: {type(retry_e).__name__}: {retry_e}",
                                usage={},
                                cost_usd=0.0,
                                metadata={"error": "retry_exception"},
                            )

            return MCPToolResult(
                content=f"Error: vLLM server returned {e.response.status_code}: {e.response.text}",
                usage={},
                cost_usd=0.0,
                metadata={"error": f"http_{e.response.status_code}"},
            )
        except Exception as e:
            return MCPToolResult(
                content=f"Error: {type(e).__name__}: {e}",
                usage={},
                cost_usd=0.0,
                metadata={"error": str(e)},
            )

    def health_check(self) -> bool:
        """Check if vLLM server is running and model is loaded."""
        try:
            with httpx.Client(timeout=5.0) as client:
                response = client.get(f"{self.vllm_url}/v1/models")
                if response.status_code == 200:
                    models = response.json().get("data", [])
                    model_ids = [m.get("id", "") for m in models]
                    return self.model_path in model_ids or self.model_name in model_ids
            return False
        except Exception:
            return False

    @classmethod
    def list_supported_models(cls) -> Dict[str, str]:
        """Return mapping of model aliases to HuggingFace paths."""
        return cls.SUPPORTED_MODELS.copy()

__init__(model_name, vllm_url='http://localhost:8000', api_key=None, telemetry_collector=None, event_recorder=None, **vllm_params)

Initialize vLLM server connection.

Parameters:

Name Type Description Default
model_name str

Model alias (e.g., 'qwen3-32b') or full HF path

 required
vllm_url str

URL of the vLLM server (default: localhost:8000)

 'http://localhost:8000'
api_key Optional[str]

Optional API key for authenticated endpoints

 None
telemetry_collector Optional[Any]

Energy monitor collector

 None
event_recorder Optional[Any]

EventRecorder for per-action tracking

 None
**vllm_params Any

Default parameters (max_tokens, temperature, top_p, etc.)

 {}
Source code in intelligence-per-watt/src/ipw/agents/mcp/vllm_server.py 
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
def __init__(
    self,
    model_name: str,
    vllm_url: str = "http://localhost:8000",
    api_key: Optional[str] = None,
    telemetry_collector: Optional[Any] = None,
    event_recorder: Optional[Any] = None,
    **vllm_params: Any,
):
    """Initialize vLLM server connection.

    Args:
        model_name: Model alias (e.g., 'qwen3-32b') or full HF path
        vllm_url: URL of the vLLM server (default: localhost:8000)
        api_key: Optional API key for authenticated endpoints
        telemetry_collector: Energy monitor collector
        event_recorder: EventRecorder for per-action tracking
        **vllm_params: Default parameters (max_tokens, temperature, top_p, etc.)
    """
    super().__init__(
        name=f"vllm:{model_name}",
        telemetry_collector=telemetry_collector,
        event_recorder=event_recorder,
    )

    self.model_name = model_name
    self.model_path = self.SUPPORTED_MODELS.get(model_name, model_name)
    self.vllm_url = vllm_url.rstrip("/")
    self.api_key = api_key or os.environ.get("VLLM_API_KEY")
    self.vllm_params = vllm_params

    # Cost estimation (per 1M tokens)
    self.cost_per_1m = self.MODEL_COSTS.get(
        model_name,
        {"prompt": 0.0, "completion": 0.0}
    )

    # Query server's actual max_model_len to handle validation
    self._server_max_model_len: Optional[int] = None

health_check()

Check if vLLM server is running and model is loaded.

Source code in intelligence-per-watt/src/ipw/agents/mcp/vllm_server.py 
315
316
317
318
319
320
321
322
323
324
325
326
def health_check(self) -> bool:
    """Check if vLLM server is running and model is loaded."""
    try:
        with httpx.Client(timeout=5.0) as client:
            response = client.get(f"{self.vllm_url}/v1/models")
            if response.status_code == 200:
                models = response.json().get("data", [])
                model_ids = [m.get("id", "") for m in models]
                return self.model_path in model_ids or self.model_name in model_ids
        return False
    except Exception:
        return False

list_supported_models() classmethod

Return mapping of model aliases to HuggingFace paths.

Source code in intelligence-per-watt/src/ipw/agents/mcp/vllm_server.py 
328
329
330
331
@classmethod
def list_supported_models(cls) -> Dict[str, str]:
    """Return mapping of model aliases to HuggingFace paths."""
    return cls.SUPPORTED_MODELS.copy()

CalculatorServer

Bases: BaseMCPServer

MCP server for mathematical calculations.

Safely evaluates mathematical expressions using AST.

Example

calc = CalculatorServer() result = calc.execute("2 + 2 * 3") print(result.content) # "8"

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_server.py 
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
class CalculatorServer(BaseMCPServer):
    """MCP server for mathematical calculations.

    Safely evaluates mathematical expressions using AST.

    Example:
        calc = CalculatorServer()
        result = calc.execute("2 + 2 * 3")
        print(result.content)  # "8"
    """

    # Safe operators for math evaluation
    OPERATORS = {
        ast.Add: operator.add,
        ast.Sub: operator.sub,
        ast.Mult: operator.mul,
        ast.Div: operator.truediv,
        ast.Pow: operator.pow,
        ast.Mod: operator.mod,
        ast.FloorDiv: operator.floordiv,
        ast.UAdd: operator.pos,
        ast.USub: operator.neg,
    }

    FUNCTIONS = {
        "sqrt": math.sqrt,
        "sin": math.sin,
        "cos": math.cos,
        "tan": math.tan,
        "log": math.log,
        "exp": math.exp,
        "abs": abs,
        "round": round,
    }

    def __init__(self, telemetry_collector: Optional[Any] = None):
        super().__init__(
            name="calculator",
            telemetry_collector=telemetry_collector,
        )

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute mathematical calculation.

        Args:
            prompt: Mathematical expression to evaluate

        Returns:
            MCPToolResult with calculated result
        """
        # Extract expression from prompt
        expression = self._extract_expression(prompt)

        try:
            result = self._safe_eval(expression)
            content = str(result)
            error = None
        except Exception as e:
            content = f"Error: {e}"
            error = str(e)

        return MCPToolResult(
            content=content,
            usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
            cost_usd=0.0,
            metadata={
                "tool": "calculator",
                "expression": expression,
                "error": error,
            },
        )

    def _extract_expression(self, prompt: str) -> str:
        """Extract mathematical expression from prompt."""
        # Look for expression in common formats
        patterns = [
            r"calculate\s+(.+)",
            r"compute\s+(.+)",
            r"evaluate\s+(.+)",
            r"what\s+is\s+(.+)",
        ]

        for pattern in patterns:
            match = re.search(pattern, prompt.lower())
            if match:
                expr = match.group(1).strip()
                # Strip trailing punctuation
                expr = expr.rstrip("?!.,;")
                return expr

        # If no pattern matches, assume entire prompt is the expression
        return prompt.strip().rstrip("?!.,;")

    def _safe_eval(self, expression: str) -> float:
        """Safely evaluate mathematical expression using AST.

        Args:
            expression: Mathematical expression string

        Returns:
            Evaluated result

        Raises:
            ValueError: If expression contains unsafe operations
        """
        # Pre-process: Convert common math notation to Python
        # Replace ^ with ** for exponentiation (careful not to replace ^^)
        expression = re.sub(r'\^(?!\^)', '**', expression)

        # Parse expression
        try:
            node = ast.parse(expression, mode="eval").body
        except SyntaxError as e:
            raise ValueError(f"Invalid expression: {e}")

        # Evaluate recursively
        return self._eval_node(node)

    def _eval_node(self, node: ast.AST) -> float:
        """Recursively evaluate AST node."""
        if isinstance(node, ast.Constant):
            return node.value
        elif isinstance(node, ast.BinOp):
            left = self._eval_node(node.left)
            right = self._eval_node(node.right)
            op = self.OPERATORS.get(type(node.op))
            if not op:
                raise ValueError(f"Unsupported operator: {node.op}")
            return op(left, right)
        elif isinstance(node, ast.UnaryOp):
            operand = self._eval_node(node.operand)
            op = self.OPERATORS.get(type(node.op))
            if not op:
                raise ValueError(f"Unsupported operator: {node.op}")
            return op(operand)
        elif isinstance(node, ast.Call):
            func_name = node.func.id if isinstance(node.func, ast.Name) else None
            if func_name not in self.FUNCTIONS:
                raise ValueError(f"Unsupported function: {func_name}")
            func = self.FUNCTIONS[func_name]
            args = [self._eval_node(arg) for arg in node.args]
            return func(*args)
        else:
            raise ValueError(f"Unsupported expression type: {type(node)}")

WebSearchServer

Bases: BaseMCPServer

MCP server for web search via Tavily API.

Tavily provides high-quality, AI-optimized search results designed for LLM consumption with structured, relevant content.

Example

search = WebSearchServer(api_key="tvly-xxx") result = search.execute("latest AI news")

Cost: ~$0.01 per search (Tavily free tier: 1000 searches/month)

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_server.py 
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
class WebSearchServer(BaseMCPServer):
    """MCP server for web search via Tavily API.

    Tavily provides high-quality, AI-optimized search results designed
    for LLM consumption with structured, relevant content.

    Example:
        search = WebSearchServer(api_key="tvly-xxx")
        result = search.execute("latest AI news")

    Cost: ~$0.01 per search (Tavily free tier: 1000 searches/month)
    """

    # Cost per search in USD
    COST_PER_SEARCH = 0.01

    def __init__(
        self,
        api_key: Optional[str] = None,
        telemetry_collector: Optional[Any] = None,
    ):
        super().__init__(
            name="web_search",
            telemetry_collector=telemetry_collector,
        )
        self.api_key = api_key or os.environ.get("TAVILY_API_KEY")
        self._client = None

    def _get_client(self):
        """Lazily initialize Tavily client."""
        if self._client is None:
            if not self.api_key:
                raise ValueError(
                    "TAVILY_API_KEY not set. Get a free API key at https://tavily.com"
                )
            from tavily import TavilyClient
            self._client = TavilyClient(api_key=self.api_key)

        return self._client

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute web search via Tavily API.

        Args:
            prompt: Search query
            **params: Additional parameters:
                - max_results: Number of results (default: 5)
                - search_depth: 'basic' or 'advanced' (default: 'basic')
                - include_answer: Include AI-generated answer (default: True)

        Returns:
            MCPToolResult with formatted search results
        """
        max_results = params.get("max_results", 5)
        search_depth = params.get("search_depth", "basic")
        include_answer = params.get("include_answer", True)

        # If no API key, return helpful message
        if not self.api_key:
            content = (
                f"[Web search for: {prompt}]\n\n"
                "Web search requires TAVILY_API_KEY environment variable.\n"
                "Get a free API key at: https://tavily.com\n"
                "Then set: export TAVILY_API_KEY='your-key'"
            )
            return MCPToolResult(
                content=content,
                usage={},
                cost_usd=0.0,
                metadata={"tool": "web_search", "error": "no_api_key"},
            )

        try:
            client = self._get_client()
            response = client.search(
                query=prompt,
                max_results=max_results,
                search_depth=search_depth,
                include_answer=include_answer,
            )

            # Format results
            lines = [f"Web search results for: {prompt}\n"]

            # Include AI-generated answer if available
            if include_answer and response.get("answer"):
                lines.append(f"Summary: {response['answer']}\n")

            # Format individual results
            results = response.get("results", [])
            for i, result in enumerate(results, 1):
                title = result.get("title", "No title")
                url = result.get("url", "")
                content_snippet = result.get("content", "")
                lines.append(f"{i}. {title}")
                lines.append(f"   URL: {url}")
                lines.append(f"   {content_snippet}")
                lines.append("")

            content = "\n".join(lines)

            return MCPToolResult(
                content=content,
                usage={},
                cost_usd=self.COST_PER_SEARCH,
                metadata={
                    "tool": "web_search",
                    "query": prompt,
                    "num_results": len(results),
                    "search_depth": search_depth,
                },
            )

        except ImportError as e:
            return MCPToolResult(
                content=f"Error: {e}",
                usage={},
                cost_usd=0.0,
                metadata={"tool": "web_search", "error": "import_error"},
            )
        except Exception as e:
            return MCPToolResult(
                content=f"Search error: {type(e).__name__}: {e}",
                usage={},
                cost_usd=0.0,
                metadata={"tool": "web_search", "error": str(e)},
            )

CodeInterpreterServer

Bases: BaseMCPServer

MCP server for Python code execution with optional sandbox isolation.

Executes Python code in a subprocess with timeout protection. Supports bubblewrap (bwrap) for filesystem isolation on Linux.

Isolation modes
  • None: Direct subprocess execution (default, for compatibility)
  • "bubblewrap": Linux namespace isolation with read-only root fs
  • "auto": Use bubblewrap if available, fall back to direct execution
Example

interpreter = CodeInterpreterServer(isolation="auto") result = interpreter.execute("print([x**2 for x in range(10)])")

Cost: ~$0.0000083 per second of compute (based on cloud GPU rates)

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_server.py 
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
class CodeInterpreterServer(BaseMCPServer):
    """MCP server for Python code execution with optional sandbox isolation.

    Executes Python code in a subprocess with timeout protection.
    Supports bubblewrap (bwrap) for filesystem isolation on Linux.

    Isolation modes:
        - None: Direct subprocess execution (default, for compatibility)
        - "bubblewrap": Linux namespace isolation with read-only root fs
        - "auto": Use bubblewrap if available, fall back to direct execution

    Example:
        interpreter = CodeInterpreterServer(isolation="auto")
        result = interpreter.execute("print([x**2 for x in range(10)])")

    Cost: ~$0.0000083 per second of compute (based on cloud GPU rates)
    """

    # Approximate cost per second of compute (GPU instance rate)
    COST_PER_SECOND = 0.0000083

    # Blocked imports for safety (used in non-isolated mode)
    BLOCKED_IMPORTS = {
        "os.system", "subprocess", "shutil.rmtree", "pathlib.Path.rmdir",
        "eval", "exec", "__import__", "importlib",
    }

    def __init__(
        self,
        timeout: int = 30,
        max_output_length: int = 10000,
        telemetry_collector: Optional[Any] = None,
        isolation: Optional[str] = None,
        allowed_paths: Optional[List[str]] = None,
    ):
        """Initialize code interpreter.

        Args:
            timeout: Maximum execution time in seconds (default: 30)
            max_output_length: Maximum characters to return (default: 10000)
            telemetry_collector: Energy monitor collector
            isolation: Isolation mode - None, "bubblewrap", or "auto"
            allowed_paths: Additional paths to mount read-only in sandbox
        """
        super().__init__(
            name="code_interpreter",
            telemetry_collector=telemetry_collector,
        )
        self.timeout = timeout
        self.max_output_length = max_output_length
        self.allowed_paths = allowed_paths or []

        # Determine isolation mode
        self._use_bubblewrap = False
        if isolation == "bubblewrap":
            if not _check_bubblewrap_available():
                raise RuntimeError(
                    "Bubblewrap (bwrap) not found. Install with: "
                    "apt-get install bubblewrap (Debian/Ubuntu) or "
                    "dnf install bubblewrap (Fedora/RHEL)"
                )
            self._use_bubblewrap = True
        elif isolation == "auto":
            self._use_bubblewrap = _check_bubblewrap_available()

        # Cache Python paths for bubblewrap
        if self._use_bubblewrap:
            self._python_paths = _get_python_lib_paths()

        self.code_extractor = re.compile(r"```[^\n]*\n([\s\S]*?)```", re.DOTALL)

    def _build_bwrap_command(self, script_path: str, sandbox_dir: str) -> List[str]:
        """Build the bubblewrap command for isolated execution."""
        cmd = [
            "bwrap",
            "--unshare-all",
            "--share-net",
            "--die-with-parent",
            "--new-session",
        ]

        # Mount system paths read-only
        for path in self._python_paths:
            if os.path.isdir(path):
                cmd.extend(["--ro-bind", path, path])
            elif os.path.isfile(path):
                cmd.extend(["--ro-bind", path, path])

        # Mount additional allowed paths read-only
        for path in self.allowed_paths:
            if os.path.exists(path):
                real_path = os.path.realpath(path)
                cmd.extend(["--ro-bind", real_path, real_path])

        # Essential virtual filesystems
        cmd.extend([
            "--proc", "/proc",
            "--dev", "/dev",
        ])

        # Create isolated /tmp with our sandbox directory
        cmd.extend([
            "--tmpfs", "/tmp",
            "--bind", sandbox_dir, "/sandbox",
            "--chdir", "/sandbox",
        ])

        # Set environment
        cmd.extend([
            "--setenv", "HOME", "/sandbox",
            "--setenv", "TMPDIR", "/tmp",
            "--setenv", "PATH", "/usr/bin:/bin:/usr/local/bin",
            "--setenv", "PYTHONDONTWRITEBYTECODE", "1",
            "--setenv", "PYTHONUNBUFFERED", "1",
        ])

        # Preserve PYTHONPATH for package access
        if os.environ.get("PYTHONPATH"):
            cmd.extend(["--setenv", "PYTHONPATH", os.environ["PYTHONPATH"]])

        # Add the Python command
        cmd.extend([sys.executable, script_path])

        return cmd

    def _preprocess_code(self, code: str) -> str:
        """Preprocess code to extract from markdown blocks."""
        match = self.code_extractor.search(code)
        if match:
            code = match.group(1).strip()
        return code

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute Python code in subprocess with optional isolation."""
        timeout = params.get("timeout", self.timeout)
        code = prompt.strip()

        # Basic safety check - warn about potentially dangerous operations
        dangerous_patterns = [
            "os.system", "subprocess.run", "subprocess.call", "subprocess.Popen",
            "shutil.rmtree", "os.remove", "os.rmdir", "__import__",
            "open(", "eval(", "exec(",
        ]
        warnings: List[str] = []

        code = self._preprocess_code(code)

        # Only warn in non-isolated mode (these are safe in sandbox)
        if not self._use_bubblewrap:
            for pattern in dangerous_patterns:
                if pattern in code:
                    warnings.append(f"Warning: Code contains '{pattern}'")

        try:
            if self._use_bubblewrap:
                return self._execute_sandboxed(code, timeout, warnings, params)
            else:
                return self._execute_direct(code, timeout, warnings, params)

        except subprocess.TimeoutExpired:
            return MCPToolResult(
                content=f"Error: Code execution timed out after {timeout} seconds",
                usage={},
                cost_usd=timeout * self.COST_PER_SECOND,
                metadata={"tool": "code_interpreter", "error": "timeout"},
            )
        except Exception as e:
            return MCPToolResult(
                content=f"Execution error: {type(e).__name__}: {e}",
                usage={},
                cost_usd=0.0,
                metadata={"tool": "code_interpreter", "error": str(e)},
            )

    def _execute_direct(
        self,
        code: str,
        timeout: int,
        warnings: List[str],
        params: dict,
    ) -> MCPToolResult:
        """Execute code directly in subprocess (no isolation)."""
        with tempfile.NamedTemporaryFile(
            mode="w",
            suffix=".py",
            delete=False
        ) as f:
            f.write(code)
            temp_file = f.name

        try:
            result = subprocess.run(
                [sys.executable, temp_file],
                capture_output=True,
                text=True,
                timeout=timeout,
                cwd=params.get("working_dir"),
            )
            return self._format_result(result, timeout, warnings, isolated=False)
        finally:
            try:
                os.unlink(temp_file)
            except OSError:
                pass

    def _execute_sandboxed(
        self,
        code: str,
        timeout: int,
        warnings: List[str],
        params: dict,
    ) -> MCPToolResult:
        """Execute code in bubblewrap sandbox with isolated filesystem."""
        with tempfile.TemporaryDirectory(prefix="ipw_sandbox_") as sandbox_dir:
            # Write script to sandbox
            script_path = os.path.join(sandbox_dir, "script.py")
            with open(script_path, "w") as f:
                f.write(code)

            # Build and run sandboxed command
            cmd = self._build_bwrap_command("/sandbox/script.py", sandbox_dir)

            # Optionally disable network
            if not params.get("network", True):
                if "--share-net" in cmd:
                    cmd.remove("--share-net")

            result = subprocess.run(
                cmd,
                capture_output=True,
                text=True,
                timeout=timeout,
            )

            return self._format_result(result, timeout, warnings, isolated=True)

    def _format_result(
        self,
        result: subprocess.CompletedProcess,
        timeout: int,
        warnings: List[str],
        isolated: bool,
    ) -> MCPToolResult:
        """Format subprocess result into MCPToolResult."""
        stdout = result.stdout
        stderr = result.stderr
        return_code = result.returncode

        # Truncate if too long
        if len(stdout) > self.max_output_length:
            stdout = stdout[:self.max_output_length] + "\n... (output truncated)"
        if len(stderr) > self.max_output_length:
            stderr = stderr[:self.max_output_length] + "\n... (output truncated)"

        # Format output
        lines = []
        if warnings:
            lines.extend(warnings)
            lines.append("")

        if stdout:
            lines.append("Output:")
            lines.append(stdout)

        if stderr:
            lines.append("Errors:")
            lines.append(stderr)

        if return_code != 0:
            lines.append(f"\nExit code: {return_code}")

        if not stdout and not stderr:
            lines.append("(No output)")

        content = "\n".join(lines)
        cost_usd = timeout * self.COST_PER_SECOND

        return MCPToolResult(
            content=content,
            usage={},
            cost_usd=cost_usd,
            metadata={
                "tool": "code_interpreter",
                "return_code": return_code,
                "timeout": timeout,
                "warnings": warnings,
                "isolated": isolated,
            },
        )

__init__(timeout=30, max_output_length=10000, telemetry_collector=None, isolation=None, allowed_paths=None)

Initialize code interpreter.

Parameters:

Name Type Description Default
timeout int

Maximum execution time in seconds (default: 30)

 30
max_output_length int

Maximum characters to return (default: 10000)

 10000
telemetry_collector Optional[Any]

Energy monitor collector

 None
isolation Optional[str]

Isolation mode - None, "bubblewrap", or "auto"

 None
allowed_paths Optional[List[str]]

Additional paths to mount read-only in sandbox

 None
Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_server.py 
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
def __init__(
    self,
    timeout: int = 30,
    max_output_length: int = 10000,
    telemetry_collector: Optional[Any] = None,
    isolation: Optional[str] = None,
    allowed_paths: Optional[List[str]] = None,
):
    """Initialize code interpreter.

    Args:
        timeout: Maximum execution time in seconds (default: 30)
        max_output_length: Maximum characters to return (default: 10000)
        telemetry_collector: Energy monitor collector
        isolation: Isolation mode - None, "bubblewrap", or "auto"
        allowed_paths: Additional paths to mount read-only in sandbox
    """
    super().__init__(
        name="code_interpreter",
        telemetry_collector=telemetry_collector,
    )
    self.timeout = timeout
    self.max_output_length = max_output_length
    self.allowed_paths = allowed_paths or []

    # Determine isolation mode
    self._use_bubblewrap = False
    if isolation == "bubblewrap":
        if not _check_bubblewrap_available():
            raise RuntimeError(
                "Bubblewrap (bwrap) not found. Install with: "
                "apt-get install bubblewrap (Debian/Ubuntu) or "
                "dnf install bubblewrap (Fedora/RHEL)"
            )
        self._use_bubblewrap = True
    elif isolation == "auto":
        self._use_bubblewrap = _check_bubblewrap_available()

    # Cache Python paths for bubblewrap
    if self._use_bubblewrap:
        self._python_paths = _get_python_lib_paths()

    self.code_extractor = re.compile(r"```[^\n]*\n([\s\S]*?)```", re.DOTALL)

ThinkServer

Bases: BaseMCPServer

MCP server for internal reasoning/scratchpad.

This is a "thinking" tool that allows the model to break down complex problems step-by-step before delegating to other tools. It simply returns the input thought process without any processing.

Example

think = ThinkServer() result = think.execute("Let me break this down: 1) First... 2) Then...")

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_server.py 
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
class ThinkServer(BaseMCPServer):
    """MCP server for internal reasoning/scratchpad.

    This is a "thinking" tool that allows the model to break down
    complex problems step-by-step before delegating to other tools.
    It simply returns the input thought process without any processing.

    Example:
        think = ThinkServer()
        result = think.execute("Let me break this down: 1) First... 2) Then...")
    """

    def __init__(self, telemetry_collector: Optional[Any] = None):
        super().__init__(
            name="think",
            telemetry_collector=telemetry_collector,
        )

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Process thinking/reasoning (pass-through)."""
        return MCPToolResult(
            content=f"[Thinking]\n{prompt}",
            usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
            cost_usd=0.0,
            metadata={"tool": "think"},
        )

FileReadServer

Bases: BaseMCPServer

MCP server for reading file contents.

Security: Only allows reading files within allowed directories.

Example

reader = FileReadServer(allowed_dirs=["/workspace"]) result = reader.execute("/workspace/file.txt")

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_server.py 
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
class FileReadServer(BaseMCPServer):
    """MCP server for reading file contents.

    Security: Only allows reading files within allowed directories.

    Example:
        reader = FileReadServer(allowed_dirs=["/workspace"])
        result = reader.execute("/workspace/file.txt")
    """

    def __init__(
        self,
        allowed_dirs: Optional[List[str]] = None,
        telemetry_collector: Optional[Any] = None,
    ):
        super().__init__(name="file_read", telemetry_collector=telemetry_collector)
        self.allowed_dirs = allowed_dirs or [os.getcwd()]

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Read file contents.

        Args:
            prompt: File path to read
            **params: start_line (1-indexed, default 1), end_line (optional)

        Returns:
            MCPToolResult with file contents or error message
        """
        file_path = prompt.strip()
        start_line = params.get("start_line", 1)
        end_line = params.get("end_line")

        # Security: resolve path and check if within allowed dirs
        try:
            resolved = os.path.realpath(file_path)
            if not any(
                resolved.startswith(os.path.realpath(d)) for d in self.allowed_dirs
            ):
                return MCPToolResult(
                    content="Error: Path not in allowed directories",
                    usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                    cost_usd=0.0,
                    metadata={"tool": "file_read", "error": "permission_denied"},
                )

            with open(resolved, "r", encoding="utf-8", errors="replace") as f:
                lines = f.readlines()

            # Apply line range (1-indexed)
            start_idx = max(0, start_line - 1)
            if end_line is not None:
                lines = lines[start_idx:end_line]
            else:
                lines = lines[start_idx:]

            content = "".join(lines)
            return MCPToolResult(
                content=content,
                usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                cost_usd=0.0,
                metadata={
                    "tool": "file_read",
                    "path": resolved,
                    "lines_read": len(lines),
                    "start_line": start_line,
                    "end_line": end_line,
                },
            )
        except FileNotFoundError:
            return MCPToolResult(
                content=f"Error: File not found: {file_path}",
                usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                cost_usd=0.0,
                metadata={"tool": "file_read", "error": "file_not_found"},
            )
        except Exception as e:
            return MCPToolResult(
                content=f"Error: {e}",
                usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                cost_usd=0.0,
                metadata={"tool": "file_read", "error": str(e)},
            )

FileWriteServer

Bases: BaseMCPServer

MCP server for writing file contents.

Security: Only allows writing files within allowed directories. Creates parent directories if they don't exist.

Example

writer = FileWriteServer(allowed_dirs=["/workspace"]) result = writer.execute("/workspace/output.txt", content="Hello, World!")

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_server.py 
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
class FileWriteServer(BaseMCPServer):
    """MCP server for writing file contents.

    Security: Only allows writing files within allowed directories.
    Creates parent directories if they don't exist.

    Example:
        writer = FileWriteServer(allowed_dirs=["/workspace"])
        result = writer.execute("/workspace/output.txt", content="Hello, World!")
    """

    def __init__(
        self,
        allowed_dirs: Optional[List[str]] = None,
        telemetry_collector: Optional[Any] = None,
    ):
        super().__init__(name="file_write", telemetry_collector=telemetry_collector)
        self.allowed_dirs = allowed_dirs or [os.getcwd()]

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Write content to file.

        Args:
            prompt: File path to write
            **params: content (required), mode ('w' for overwrite or 'a' for append)

        Returns:
            MCPToolResult with success message or error
        """
        file_path = prompt.strip()
        content = params.get("content", "")
        mode = params.get("mode", "w")

        # Validate mode
        if mode not in ("w", "a"):
            return MCPToolResult(
                content=f"Error: Invalid mode '{mode}'. Use 'w' (write) or 'a' (append).",
                usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                cost_usd=0.0,
                metadata={"tool": "file_write", "error": "invalid_mode"},
            )

        # Security: resolve path and check if within allowed dirs
        try:
            # For new files, resolve the parent directory
            parent_dir = os.path.dirname(file_path)
            if parent_dir:
                os.makedirs(parent_dir, exist_ok=True)

            resolved = os.path.realpath(file_path)
            if not any(
                resolved.startswith(os.path.realpath(d)) for d in self.allowed_dirs
            ):
                return MCPToolResult(
                    content="Error: Path not in allowed directories",
                    usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                    cost_usd=0.0,
                    metadata={"tool": "file_write", "error": "permission_denied"},
                )

            with open(resolved, mode, encoding="utf-8") as f:
                f.write(content)

            action = "appended to" if mode == "a" else "wrote"
            return MCPToolResult(
                content=f"Successfully {action} {len(content)} chars to {resolved}",
                usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                cost_usd=0.0,
                metadata={
                    "tool": "file_write",
                    "path": resolved,
                    "bytes_written": len(content),
                    "mode": mode,
                },
            )
        except Exception as e:
            return MCPToolResult(
                content=f"Error: {e}",
                usage={"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0},
                cost_usd=0.0,
                metadata={"tool": "file_write", "error": str(e)},
            )

ToolRegistry

Unified registry for all ToolOrchestra + ADP tools.

Example

registry = ToolRegistry() registry.discover_tools()

tools = registry.get_available_tools() small_llms = registry.get_tools_by_category(ToolCategory.LLM_SMALL)

calc = registry.get_tool_instance("calculator") result = calc.execute("2 + 2")

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
class ToolRegistry:
    """Unified registry for all ToolOrchestra + ADP tools.

    Example:
        registry = ToolRegistry()
        registry.discover_tools()

        tools = registry.get_available_tools()
        small_llms = registry.get_tools_by_category(ToolCategory.LLM_SMALL)

        calc = registry.get_tool_instance("calculator")
        result = calc.execute("2 + 2")
    """

    def __init__(
        self,
        ollama_base_url: str = "http://localhost:11434",
        vllm_base_url: str = "http://localhost:8000",
        telemetry_collector: Optional[Any] = None,
        code_isolation: Optional[str] = "auto",
        retrieval_gpu_device: Optional[int] = None,
    ):
        """Initialize tool registry.

        Args:
            ollama_base_url: Base URL for Ollama server
            vllm_base_url: Base URL for vLLM server
            telemetry_collector: Energy monitor collector for all tools
            code_isolation: Isolation mode for code_interpreter tool.
            retrieval_gpu_device: GPU device index for neural retrieval models.
        """
        self.ollama_base_url = ollama_base_url
        self.vllm_base_url = vllm_base_url
        self.telemetry_collector = telemetry_collector
        self.code_isolation = code_isolation
        self.retrieval_gpu_device = retrieval_gpu_device

        self._specs: Dict[str, ToolSpec] = {}
        self._instances: Dict[str, BaseMCPServer] = {}
        self._aliases: Dict[str, str] = {}

        # Register all known tools
        self._register_builtin_tools()

    def _register_builtin_tools(self):
        """Register all built-in tools from ToolOrchestra + ADP."""

        # === UTILITY TOOLS ===
        self.register(ToolSpec(
            name="calculator",
            category=ToolCategory.UTILITY,
            description="Evaluate mathematical expressions. Supports arithmetic, exponents, trig functions. Zero cost, instant.",
            estimated_latency_ms=1,
            estimated_cost_usd=0.0,
            capabilities=["math", "arithmetic", "computation"],
            adp_domains=["agenttuning_db", "agenttuning_kg"],
        ))

        self.register(ToolSpec(
            name="think",
            category=ToolCategory.UTILITY,
            description="Internal reasoning scratchpad. Break down complex problems step-by-step before delegating. Zero cost.",
            estimated_latency_ms=1,
            estimated_cost_usd=0.0,
            capabilities=["reasoning", "planning", "decomposition"],
            adp_domains=["codeact", "agenttuning_alfworld"],
        ))

        # === CODE TOOLS ===
        self.register(ToolSpec(
            name="code_interpreter",
            category=ToolCategory.CODE,
            description="Execute Python code in sandbox. Returns stdout/stderr. 30s timeout.",
            estimated_latency_ms=1000,
            estimated_cost_usd=0.00001,
            capabilities=["code_execution", "python", "computation"],
            adp_domains=["codeact", "code_feedback", "swe-smith"],
        ))

        self.register(ToolSpec(
            name="file_read",
            category=ToolCategory.CODE,
            description="Read file contents. Supports line ranges. Zero cost, instant.",
            estimated_latency_ms=10,
            estimated_cost_usd=0.0,
            capabilities=["file_operations", "code_analysis"],
            adp_domains=["codeact", "swe-smith"],
        ))

        self.register(ToolSpec(
            name="file_write",
            category=ToolCategory.CODE,
            description="Write content to file. Supports write and append modes. Zero cost, instant.",
            estimated_latency_ms=10,
            estimated_cost_usd=0.0,
            capabilities=["file_operations", "code_generation"],
            adp_domains=["codeact", "swe-smith"],
        ))

        # === SEARCH TOOLS ===
        self.register(ToolSpec(
            name="web_search",
            category=ToolCategory.SEARCH,
            description="Search the web via Tavily API. Cost: $0.01/search.",
            estimated_latency_ms=500,
            estimated_cost_usd=0.01,
            requires_api_key="TAVILY_API_KEY",
            capabilities=["search", "retrieval", "current_info"],
            adp_domains=["agenttuning_webshop", "mind2web", "go-browse-wa"],
        ))

        # === RETRIEVAL TOOLS ===
        self.register(ToolSpec(
            name="retrieval:grep",
            category=ToolCategory.SEARCH,
            description="Fast regex/keyword search. No indexing, ~1ms latency.",
            estimated_latency_ms=1,
            estimated_cost_usd=0.0,
            capabilities=["search", "retrieval", "keyword_search", "regex"],
        ))

        self.register(ToolSpec(
            name="retrieval:bm25",
            category=ToolCategory.SEARCH,
            description="BM25 sparse retrieval. Fast, CPU-only, ~10ms latency.",
            estimated_latency_ms=10,
            estimated_cost_usd=0.0,
            capabilities=["search", "retrieval", "keyword_search"],
        ))

        self.register(ToolSpec(
            name="retrieval:dense",
            category=ToolCategory.SEARCH,
            description="Dense neural retrieval with FAISS. Semantic search, ~50ms.",
            estimated_latency_ms=50,
            estimated_cost_usd=0.0,
            capabilities=["search", "retrieval", "semantic_search"],
        ))

        self.register(ToolSpec(
            name="retrieval:hybrid",
            category=ToolCategory.SEARCH,
            description="Hybrid BM25 + dense retrieval with RRF fusion. Best accuracy, ~100ms.",
            estimated_latency_ms=100,
            estimated_cost_usd=0.0,
            capabilities=["search", "retrieval", "semantic_search", "keyword_search"],
        ))

        # === SMALL LLMs (<3B) via Ollama ===
        small_llms = [
            ("ollama:qwen2.5:0.5b", "Qwen2.5 0.5B - Fastest, basic tasks", 300, ["basic_qa"]),
            ("ollama:qwen2.5:1.5b", "Qwen2.5 1.5B - Fast, simple reasoning", 800, ["simple_reasoning"]),
            ("ollama:qwen3:1.5b", "Qwen3 1.5B - Fast reasoning with Qwen3 architecture", 800, ["simple_reasoning"]),
            ("ollama:llama3.2:1b", "Llama3.2 1B - Fast, general tasks", 500, ["basic_qa"]),
        ]
        for name, desc, latency, caps in small_llms:
            self.register(ToolSpec(
                name=name,
                category=ToolCategory.LLM_SMALL,
                description=desc,
                estimated_latency_ms=latency,
                estimated_cost_usd=0.0,
                requires_server="ollama",
                capabilities=caps + ["text_generation"],
            ))

        # === MEDIUM LLMs (3-10B) via Ollama ===
        medium_llms = [
            ("ollama:qwen2.5:3b", "Qwen2.5 3B - Balanced speed/quality", 1500, ["reasoning"]),
            ("ollama:qwen2.5:7b", "Qwen2.5 7B - Good reasoning", 3000, ["reasoning", "complex_qa"]),
            ("ollama:llama3.2:3b", "Llama3.2 3B - Balanced, general", 1500, ["reasoning"]),
            ("vllm:qwen3-8b", "Qwen3 8B - High quality reasoning", 2000, ["complex_reasoning"]),
            ("vllm:llama-8b", "Llama3.1 8B - Strong general model", 2000, ["complex_reasoning"]),
        ]
        for name, desc, latency, caps in medium_llms:
            self.register(ToolSpec(
                name=name,
                category=ToolCategory.LLM_MEDIUM,
                description=desc,
                estimated_latency_ms=latency,
                estimated_cost_usd=0.0,
                requires_server="ollama",
                capabilities=caps + ["text_generation"],
            ))

        # === LARGE LLMs (>10B) via vLLM ===
        large_llms = [
            ("vllm:qwen3-32b", "Qwen3 32B - Best open-source quality", 5000, ["complex_reasoning", "analysis"]),
            ("vllm:llama-70b", "Llama3.1 70B - Near SOTA quality", 8000, ["complex_reasoning", "analysis"]),
        ]
        for name, desc, latency, caps in large_llms:
            self.register(ToolSpec(
                name=name,
                category=ToolCategory.LLM_LARGE,
                description=desc,
                estimated_latency_ms=latency,
                estimated_cost_usd=0.0,
                requires_server="vllm",
                capabilities=caps + ["text_generation"],
            ))

        # === SPECIALIST LLMs via vLLM ===
        specialist_llms = [
            ("vllm:qwen-math-7b", "Qwen Math 7B - Math specialist", 2000, ["math", "problem_solving"]),
            ("vllm:glm-4.7", "GLM-4.7 - Best math model", 8000, ["math", "complex_math"]),
            ("vllm:qwen-coder-7b", "Qwen Coder 7B - Code specialist", 2000, ["code", "programming"]),
            ("vllm:qwen3-coder-plus", "Qwen3 Coder Plus - Best code model", 5000, ["code", "complex_code"]),
        ]
        for name, desc, latency, caps in specialist_llms:
            self.register(ToolSpec(
                name=name,
                category=ToolCategory.LLM_SPECIALIST,
                description=desc,
                estimated_latency_ms=latency,
                estimated_cost_usd=0.0,
                requires_server="vllm",
                capabilities=caps + ["text_generation"],
                adp_domains=["codeact", "code_feedback"] if "code" in caps else [],
            ))

        # === LLM ALIASES ===
        llm_aliases = [
            ("llm_small", "vllm:qwen3-1.5b", "Small LLM for fast reasoning (Qwen3 1.5B)",
             ToolCategory.LLM_SMALL, 100, ["simple_reasoning"]),
            ("llm_medium", "vllm:qwen3-8b", "Medium LLM for balanced tasks (Qwen3 8B)",
             ToolCategory.LLM_MEDIUM, 300, ["reasoning"]),
            ("llm_large", "vllm:qwen3-32b", "Large LLM for complex reasoning (Qwen3 32B)",
             ToolCategory.LLM_LARGE, 1000, ["complex_reasoning"]),
            ("llm_specialist", "vllm:qwen-coder-32b", "Specialist LLM for code (Qwen Coder 32B)",
             ToolCategory.LLM_SPECIALIST, 1000, ["code", "programming"]),
        ]
        for alias_name, target, desc, category, latency, caps in llm_aliases:
            self.register(ToolSpec(
                name=alias_name,
                category=category,
                description=desc,
                estimated_latency_ms=latency,
                estimated_cost_usd=0.0,
                requires_server="vllm",
                capabilities=caps + ["text_generation"],
            ))
            self._aliases[alias_name] = target

        # === CLOUD LLMs ===
        cloud_llms = [
            ("openai:gpt-5-mini-2025-08-07", "GPT-5 Mini - Fast, capable cloud", 800, 0.005, ["reasoning"], "OPENAI_API_KEY"),
            ("openai:gpt-4o", "GPT-4o - Best GPT-4 model", 1000, 0.0025, ["complex_reasoning"], "OPENAI_API_KEY"),
            ("openai:o1-mini", "o1-mini - Reasoning model", 2000, 0.003, ["deep_reasoning"], "OPENAI_API_KEY"),
            ("openai:o1", "o1 - Best reasoning model", 5000, 0.015, ["deep_reasoning"], "OPENAI_API_KEY"),
            ("openai:gpt-5.2-2025-12-11", "GPT-5.2 - Most capable OpenAI model", 2000, 0.03, ["complex_reasoning", "analysis"], "OPENAI_API_KEY"),
            ("openai:gpt-5-nano-2025-08-07", "GPT-5 Nano - Fastest, cheapest", 400, 0.001, ["basic_reasoning", "fast"], "OPENAI_API_KEY"),
            ("anthropic:claude-3-5-haiku-20241022", "Claude 3.5 Haiku - Fast, cheap", 400, 0.0008, ["reasoning"], "ANTHROPIC_API_KEY"),
            ("anthropic:claude-sonnet-4-20250514", "Claude Sonnet 4 - Balanced", 800, 0.003, ["complex_reasoning"], "ANTHROPIC_API_KEY"),
            ("anthropic:claude-opus-4-20250514", "Claude Opus 4 - Most capable", 2000, 0.015, ["complex_reasoning", "analysis"], "ANTHROPIC_API_KEY"),
            ("anthropic:claude-haiku-4-5-20251001", "Claude 4.5 Haiku - Fast, cheap", 300, 0.001, ["reasoning", "fast"], "ANTHROPIC_API_KEY"),
            ("anthropic:claude-sonnet-4-5-20250929", "Claude 4.5 Sonnet - Balanced quality/speed", 600, 0.004, ["complex_reasoning"], "ANTHROPIC_API_KEY"),
            ("anthropic:claude-opus-4-5-20251101", "Claude 4.5 Opus - Most capable Anthropic model", 1500, 0.02, ["complex_reasoning", "analysis", "deep_reasoning"], "ANTHROPIC_API_KEY"),
        ]
        for name, desc, latency, cost, caps, api_key in cloud_llms:
            self.register(ToolSpec(
                name=name,
                category=ToolCategory.LLM_CLOUD,
                description=desc,
                estimated_latency_ms=latency,
                estimated_cost_usd=cost,
                requires_api_key=api_key,
                capabilities=caps + ["text_generation"],
            ))

        # === OPENROUTER ===
        openrouter_models = [
            ("openrouter:google/gemini-2.5-flash", "Gemini 2.5 Flash via OpenRouter", 500, 0.00015, ["reasoning", "fast"]),
            ("openrouter:google/gemini-2.5-pro", "Gemini 2.5 Pro via OpenRouter", 1000, 0.00125, ["complex_reasoning"]),
            ("openrouter:anthropic/claude-sonnet-4", "Claude Sonnet 4 via OpenRouter", 800, 0.003, ["complex_reasoning"]),
            ("openrouter:openai/gpt-4o", "GPT-4o via OpenRouter", 1000, 0.0025, ["complex_reasoning"]),
            ("openrouter:openai/gpt-5-mini-2025-08-07", "GPT-5 Mini via OpenRouter", 800, 0.005, ["reasoning"]),
            ("openrouter:meta-llama/llama-3.3-70b-instruct", "Llama 3.3 70B via OpenRouter", 2000, 0.0004, ["reasoning"]),
            ("openrouter:qwen/qwen-2.5-72b-instruct", "Qwen 2.5 72B via OpenRouter", 2000, 0.00035, ["reasoning"]),
            ("openrouter:qwen/qwq-32b", "QwQ 32B reasoning model via OpenRouter", 3000, 0.00015, ["deep_reasoning"]),
            ("openrouter:deepseek/deepseek-r1", "DeepSeek R1 reasoning via OpenRouter", 3000, 0.00055, ["deep_reasoning"]),
            ("openrouter:deepseek/deepseek-chat-v3-0324", "DeepSeek Chat V3 via OpenRouter", 1000, 0.00014, ["reasoning"]),
            ("openrouter:mistralai/mistral-large-2411", "Mistral Large via OpenRouter", 1500, 0.002, ["reasoning"]),
            ("openrouter:qwen/qwen3-32b", "Qwen3 32B via OpenRouter", 3000, 0.0002, ["complex_reasoning", "analysis"]),
            ("openrouter:z-ai/glm-4.7", "GLM-4.7 via OpenRouter - Best math model", 4000, 0.0004, ["math", "complex_math", "problem_solving"]),
            ("openrouter:qwen/qwen3-coder-plus", "Qwen3 Coder Plus via OpenRouter", 3000, 0.0002, ["code", "complex_code", "programming"]),
        ]
        for name, desc, latency, cost, caps in openrouter_models:
            self.register(ToolSpec(
                name=name,
                category=ToolCategory.LLM_CLOUD,
                description=desc,
                estimated_latency_ms=latency,
                estimated_cost_usd=cost,
                requires_api_key="OPENROUTER_API_KEY",
                capabilities=caps + ["text_generation"],
            ))

        # === ADP DOMAIN-SPECIFIC TOOLS ===
        self.register(ToolSpec(
            name="adp:codeact",
            category=ToolCategory.ADP_CODEACT,
            description="Execute code actions from ADP codeact domain.",
            capabilities=["code_execution", "reasoning"],
            adp_domains=["codeact"],
        ))

        self.register(ToolSpec(
            name="adp:alfworld",
            category=ToolCategory.ADP_ALFWORLD,
            description="Household task actions from ALFWorld domain.",
            capabilities=["embodied_actions", "planning"],
            adp_domains=["agenttuning_alfworld"],
        ))

        self.register(ToolSpec(
            name="adp:mind2web",
            category=ToolCategory.ADP_MIND2WEB,
            description="Web navigation actions from Mind2Web domain.",
            capabilities=["web_navigation", "ui_interaction"],
            adp_domains=["mind2web", "agenttuning_mind2web"],
        ))

        self.register(ToolSpec(
            name="adp:database",
            category=ToolCategory.ADP_DATABASE,
            description="Database query actions from ADP database domain.",
            capabilities=["sql", "query"],
            adp_domains=["agenttuning_db"],
        ))

    def register(self, spec: ToolSpec):
        """Register a tool specification."""
        self._specs[spec.name] = spec

    def get_spec(self, name: str) -> Optional[ToolSpec]:
        """Get tool specification by name."""
        return self._specs.get(name)

    def get_all_specs(self) -> List[ToolSpec]:
        """Get all registered tool specifications."""
        return list(self._specs.values())

    def get_specs_by_category(self, category: ToolCategory) -> List[ToolSpec]:
        """Get tool specifications by category."""
        return [s for s in self._specs.values() if s.category == category]

    def get_specs_for_domain(self, domain: str) -> List[ToolSpec]:
        """Get tool specifications relevant for an ADP domain."""
        return [s for s in self._specs.values() if domain in s.adp_domains]

    def get_specs_by_capability(self, capability: str) -> List[ToolSpec]:
        """Get tool specifications by capability tag."""
        return [s for s in self._specs.values() if capability in s.capabilities]

    def discover_available_tools(self) -> List[str]:
        """Discover which tools are actually available."""
        available = []

        for name, spec in self._specs.items():
            if spec.requires_api_key:
                if not os.environ.get(spec.requires_api_key):
                    continue
            available.append(name)

        return available

    def get_tool_instance(self, name: str) -> Optional[BaseMCPServer]:
        """Get or create a tool instance."""
        if name in self._instances:
            return self._instances[name]

        spec = self._specs.get(name)
        if not spec:
            return None

        instance = self._create_instance(name, spec)
        if instance:
            self._instances[name] = instance

        return instance

    def _create_instance(self, name: str, spec: ToolSpec) -> Optional[BaseMCPServer]:
        """Create a tool instance from specification."""
        try:
            # Check for alias first
            if name in self._aliases:
                target_name = self._aliases[name]
                target_spec = self._specs.get(target_name)
                return self._create_instance(target_name, target_spec)

            # Utility tools
            if name == "calculator":
                from ipw.agents.mcp.tool_server import CalculatorServer
                return CalculatorServer(telemetry_collector=self.telemetry_collector)

            elif name == "think":
                from ipw.agents.mcp.tool_server import ThinkServer
                return ThinkServer(telemetry_collector=self.telemetry_collector)

            elif name == "code_interpreter":
                from ipw.agents.mcp.tool_server import CodeInterpreterServer
                return CodeInterpreterServer(
                    telemetry_collector=self.telemetry_collector,
                    isolation=self.code_isolation,
                )

            elif name == "web_search":
                from ipw.agents.mcp.tool_server import WebSearchServer
                return WebSearchServer(telemetry_collector=self.telemetry_collector)

            elif name == "file_read":
                from ipw.agents.mcp.tool_server import FileReadServer
                return FileReadServer(telemetry_collector=self.telemetry_collector)

            elif name == "file_write":
                from ipw.agents.mcp.tool_server import FileWriteServer
                return FileWriteServer(telemetry_collector=self.telemetry_collector)

            # Ollama models
            elif name.startswith("ollama:"):
                from ipw.agents.mcp.ollama_server import OllamaMCPServer
                model_name = name.split(":", 1)[1]
                return OllamaMCPServer(
                    model_name=model_name,
                    base_url=self.ollama_base_url,
                    telemetry_collector=self.telemetry_collector,
                )

            # vLLM models
            elif name.startswith("vllm:"):
                from ipw.agents.mcp.vllm_server import VLLMMCPServer
                model_name = name.split(":", 1)[1]
                return VLLMMCPServer(
                    model_name=model_name,
                    vllm_url=self.vllm_base_url,
                    telemetry_collector=self.telemetry_collector,
                )

            # OpenAI models
            elif name.startswith("openai:"):
                from ipw.agents.mcp.openai_server import OpenAIMCPServer
                model_name = name.split(":", 1)[1]
                return OpenAIMCPServer(
                    model_name=model_name,
                    telemetry_collector=self.telemetry_collector,
                )

            # Anthropic models
            elif name.startswith("anthropic:"):
                from ipw.agents.mcp.anthropic_server import AnthropicMCPServer
                model_name = name.split(":", 1)[1]
                return AnthropicMCPServer(
                    model_name=model_name,
                    telemetry_collector=self.telemetry_collector,
                )

            # OpenRouter models
            elif name.startswith("openrouter:"):
                from ipw.agents.mcp.openrouter_server import OpenRouterMCPServer
                model_name = name.split(":", 1)[1]
                return OpenRouterMCPServer(
                    model_name=model_name,
                    telemetry_collector=self.telemetry_collector,
                )

            # ADP domain tools
            elif name.startswith("adp:"):
                return ADPDomainServer(
                    domain=name.split(":", 1)[1],
                    telemetry_collector=self.telemetry_collector,
                )

            # Retrieval tools
            elif name.startswith("retrieval:"):
                retrieval_type = name.split(":", 1)[1]
                if retrieval_type == "grep":
                    from ipw.agents.mcp.retrieval import GrepRetrievalServer
                    return GrepRetrievalServer(
                        telemetry_collector=self.telemetry_collector,
                    )
                elif retrieval_type == "bm25":
                    from ipw.agents.mcp.retrieval import BM25RetrievalServer
                    return BM25RetrievalServer(
                        telemetry_collector=self.telemetry_collector,
                    )
                elif retrieval_type == "dense":
                    from ipw.agents.mcp.retrieval import DenseRetrievalServer
                    use_gpu = self.retrieval_gpu_device is not None
                    return DenseRetrievalServer(
                        telemetry_collector=self.telemetry_collector,
                        use_gpu=use_gpu,
                        gpu_device=self.retrieval_gpu_device or 0,
                    )
                elif retrieval_type == "hybrid":
                    from ipw.agents.mcp.retrieval import HybridRetrievalServer
                    use_gpu = self.retrieval_gpu_device is not None
                    return HybridRetrievalServer(
                        model_name="Qwen/Qwen3-Embedding-4B",
                        telemetry_collector=self.telemetry_collector,
                        use_gpu=use_gpu,
                        gpu_device=self.retrieval_gpu_device or 0,
                    )

        except ImportError as e:
            print(f"Warning: Could not import server for '{name}': {e}")
        except Exception as e:
            print(f"Warning: Could not create instance for '{name}': {e}")

        return None

    def get_tool_descriptions(self, tools: Optional[List[str]] = None) -> str:
        """Get formatted tool descriptions for prompting."""
        if tools is None:
            tools = self.discover_available_tools()

        lines = ["Available tools:"]
        for name in tools:
            spec = self._specs.get(name)
            if spec:
                cost_info = f"${spec.estimated_cost_usd:.4f}" if spec.estimated_cost_usd > 0 else "free"
                lines.append(f"- {name}: {spec.description} ({cost_info}, ~{spec.estimated_latency_ms}ms)")

        return "\n".join(lines)

__init__(ollama_base_url='http://localhost:11434', vllm_base_url='http://localhost:8000', telemetry_collector=None, code_isolation='auto', retrieval_gpu_device=None)

Initialize tool registry.

Parameters:

Name Type Description Default
ollama_base_url str

Base URL for Ollama server

 'http://localhost:11434'
vllm_base_url str

Base URL for vLLM server

 'http://localhost:8000'
telemetry_collector Optional[Any]

Energy monitor collector for all tools

 None
code_isolation Optional[str]

Isolation mode for code_interpreter tool.

 'auto'
retrieval_gpu_device Optional[int]

GPU device index for neural retrieval models.

 None
Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
def __init__(
    self,
    ollama_base_url: str = "http://localhost:11434",
    vllm_base_url: str = "http://localhost:8000",
    telemetry_collector: Optional[Any] = None,
    code_isolation: Optional[str] = "auto",
    retrieval_gpu_device: Optional[int] = None,
):
    """Initialize tool registry.

    Args:
        ollama_base_url: Base URL for Ollama server
        vllm_base_url: Base URL for vLLM server
        telemetry_collector: Energy monitor collector for all tools
        code_isolation: Isolation mode for code_interpreter tool.
        retrieval_gpu_device: GPU device index for neural retrieval models.
    """
    self.ollama_base_url = ollama_base_url
    self.vllm_base_url = vllm_base_url
    self.telemetry_collector = telemetry_collector
    self.code_isolation = code_isolation
    self.retrieval_gpu_device = retrieval_gpu_device

    self._specs: Dict[str, ToolSpec] = {}
    self._instances: Dict[str, BaseMCPServer] = {}
    self._aliases: Dict[str, str] = {}

    # Register all known tools
    self._register_builtin_tools()

register(spec)

Register a tool specification.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
416
417
418
def register(self, spec: ToolSpec):
    """Register a tool specification."""
    self._specs[spec.name] = spec

get_spec(name)

Get tool specification by name.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
420
421
422
def get_spec(self, name: str) -> Optional[ToolSpec]:
    """Get tool specification by name."""
    return self._specs.get(name)

get_all_specs()

Get all registered tool specifications.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
424
425
426
def get_all_specs(self) -> List[ToolSpec]:
    """Get all registered tool specifications."""
    return list(self._specs.values())

get_specs_by_category(category)

Get tool specifications by category.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
428
429
430
def get_specs_by_category(self, category: ToolCategory) -> List[ToolSpec]:
    """Get tool specifications by category."""
    return [s for s in self._specs.values() if s.category == category]

get_specs_for_domain(domain)

Get tool specifications relevant for an ADP domain.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
432
433
434
def get_specs_for_domain(self, domain: str) -> List[ToolSpec]:
    """Get tool specifications relevant for an ADP domain."""
    return [s for s in self._specs.values() if domain in s.adp_domains]

get_specs_by_capability(capability)

Get tool specifications by capability tag.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
436
437
438
def get_specs_by_capability(self, capability: str) -> List[ToolSpec]:
    """Get tool specifications by capability tag."""
    return [s for s in self._specs.values() if capability in s.capabilities]

discover_available_tools()

Discover which tools are actually available.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
440
441
442
443
444
445
446
447
448
449
450
def discover_available_tools(self) -> List[str]:
    """Discover which tools are actually available."""
    available = []

    for name, spec in self._specs.items():
        if spec.requires_api_key:
            if not os.environ.get(spec.requires_api_key):
                continue
        available.append(name)

    return available

get_tool_instance(name)

Get or create a tool instance.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
452
453
454
455
456
457
458
459
460
461
462
463
464
465
def get_tool_instance(self, name: str) -> Optional[BaseMCPServer]:
    """Get or create a tool instance."""
    if name in self._instances:
        return self._instances[name]

    spec = self._specs.get(name)
    if not spec:
        return None

    instance = self._create_instance(name, spec)
    if instance:
        self._instances[name] = instance

    return instance

get_tool_descriptions(tools=None)

Get formatted tool descriptions for prompting.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
596
597
598
599
600
601
602
603
604
605
606
607
608
def get_tool_descriptions(self, tools: Optional[List[str]] = None) -> str:
    """Get formatted tool descriptions for prompting."""
    if tools is None:
        tools = self.discover_available_tools()

    lines = ["Available tools:"]
    for name in tools:
        spec = self._specs.get(name)
        if spec:
            cost_info = f"${spec.estimated_cost_usd:.4f}" if spec.estimated_cost_usd > 0 else "free"
            lines.append(f"- {name}: {spec.description} ({cost_info}, ~{spec.estimated_latency_ms}ms)")

    return "\n".join(lines)

ToolSpec dataclass

Specification for a tool in the registry.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
@dataclass
class ToolSpec:
    """Specification for a tool in the registry."""

    name: str
    """Unique tool identifier (e.g., 'calculator', 'ollama:qwen2.5:1.5b')"""

    category: ToolCategory
    """Tool category for routing decisions"""

    description: str
    """Human-readable description for policy model"""

    server_class: Optional[Type[BaseMCPServer]] = None
    """MCP server class to instantiate"""

    factory: Optional[Callable[..., BaseMCPServer]] = None
    """Factory function for custom initialization"""

    # Cost/efficiency metadata (for routing decisions)
    estimated_latency_ms: float = 0.0
    """Estimated latency in milliseconds"""

    estimated_cost_usd: float = 0.0
    """Estimated cost per call in USD"""

    estimated_energy_joules: float = 0.0
    """Estimated energy consumption per call"""

    requires_api_key: Optional[str] = None
    """Environment variable name for required API key"""

    requires_server: Optional[str] = None
    """Required server (e.g., 'ollama', 'vllm')"""

    # ADP domain mapping
    adp_domains: List[str] = field(default_factory=list)
    """ADP domains this tool is relevant for"""

    # Capability tags for semantic matching
    capabilities: List[str] = field(default_factory=list)
    """Capability tags (e.g., 'math', 'code', 'reasoning')"""

name instance-attribute

Unique tool identifier (e.g., 'calculator', 'ollama:qwen2.5:1.5b')

category instance-attribute

Tool category for routing decisions

description instance-attribute

Human-readable description for policy model

server_class = None class-attribute instance-attribute

MCP server class to instantiate

factory = None class-attribute instance-attribute

Factory function for custom initialization

estimated_latency_ms = 0.0 class-attribute instance-attribute

Estimated latency in milliseconds

estimated_cost_usd = 0.0 class-attribute instance-attribute

Estimated cost per call in USD

estimated_energy_joules = 0.0 class-attribute instance-attribute

Estimated energy consumption per call

requires_api_key = None class-attribute instance-attribute

Environment variable name for required API key

requires_server = None class-attribute instance-attribute

Required server (e.g., 'ollama', 'vllm')

adp_domains = field(default_factory=list) class-attribute instance-attribute

ADP domains this tool is relevant for

capabilities = field(default_factory=list) class-attribute instance-attribute

Capability tags (e.g., 'math', 'code', 'reasoning')

ToolCategory

Bases: Enum

Categories of tools matching ToolOrchestra + ADP domains.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
class ToolCategory(Enum):
    """Categories of tools matching ToolOrchestra + ADP domains."""

    # ToolOrchestra core tools
    UTILITY = "utility"
    CODE = "code"
    SEARCH = "search"

    # LLM backends (by size/capability)
    LLM_SMALL = "llm_small"
    LLM_MEDIUM = "llm_medium"
    LLM_LARGE = "llm_large"
    LLM_SPECIALIST = "llm_specialist"
    LLM_CLOUD = "llm_cloud"

    # ADP domain-specific actions
    ADP_CODEACT = "adp_codeact"
    ADP_ALFWORLD = "adp_alfworld"
    ADP_MIND2WEB = "adp_mind2web"
    ADP_DATABASE = "adp_database"

ADPDomainServer

Bases: BaseMCPServer

Passthrough server for ADP domain-specific actions.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
class ADPDomainServer(BaseMCPServer):
    """Passthrough server for ADP domain-specific actions."""

    def __init__(
        self,
        domain: str,
        telemetry_collector: Optional[Any] = None,
    ):
        super().__init__(
            name=f"adp:{domain}",
            telemetry_collector=telemetry_collector,
        )
        self.domain = domain

    def _execute_impl(self, prompt: str, **params: Any) -> MCPToolResult:
        """Execute ADP domain action (passthrough for training)."""
        return MCPToolResult(
            content=f"[ADP:{self.domain}] Action executed: {prompt[:100]}...",
            usage={},
            cost_usd=0.0,
            metadata={
                "tool": f"adp:{self.domain}",
                "domain": self.domain,
                "action": prompt,
            },
        )

get_registry(**kwargs)

Get or create the global tool registry.

Source code in intelligence-per-watt/src/ipw/agents/mcp/tool_registry.py 
644
645
646
647
648
649
650
651
652
653
654
655
def get_registry(**kwargs) -> ToolRegistry:
    """Get or create the global tool registry."""
    global _global_registry, _global_registry_kwargs

    if _global_registry is not None and kwargs != _global_registry_kwargs:
        _global_registry = None

    if _global_registry is None:
        _global_registry = ToolRegistry(**kwargs)
        _global_registry_kwargs = kwargs.copy()

    return _global_registry