LLM Bridge MCP

Instructions

Implementation Reference

• llm_bridge_mcp/main.py:25-25 (registration)

  Registration of the 'run_llm' tool using the @mcp.tool() decorator.

  @mcp.tool()

• llm_bridge_mcp/main.py:26-63 (handler)

  The main handler function for the 'run_llm' tool. It creates an Agent from pydantic_ai, runs the prompt with specified settings, constructs an LLMResponse, and returns its JSON dump.

  async def run_llm(
      prompt: str,
      model_name: KnownModelName = Field(
          default="openai:gpt-4o-mini",
          description=f"Specific model name. Available models: {', '.join(KnownModelName.__args__)}",
      ),
      temperature: float = Field(
          default=0.7,
          description="Controls randomness (0.0 to 1.0)",
      ),
      max_tokens: int = Field(
          default=8192,
          description="Maximum number of tokens to generate",
      ),
      system_prompt: str = Field(
          default="",
          description="Optional system prompt to guide the model's behavior",
      ),
  ) -> str:
      """Run a prompt through an LLM and return the response."""
      agent = Agent(
          model=model_name,
          system_prompt=system_prompt,
      )
      response = await agent.run(
          prompt,
          model_settings={
              "temperature": temperature,
              "max_tokens": max_tokens,
          },
      )
      res = LLMResponse(
          content=response.data,
          model_name=model_name,
          usage=response.usage(),
          temperature=temperature,
      )
      return res.model_dump_json()

• llm_bridge_mcp/main.py:16-23 (schema)

  Pydantic schema/model for the LLM response used internally by the 'run_llm' handler.

  class LLMResponse(BaseModel):
      """Response from an LLM."""
  
      content: str
      model_name: str
      usage: Usage
      temperature: float