MCP DeepInfra AI Tools Server

Instructions

Implementation Reference

• src/mcp_deepinfra/server.py:266-286 (handler)

  The handler function that implements the fill_mask tool logic. It constructs a prompt to fill in masked text using the completions API with a specified model and returns the generated response.

      async def fill_mask(text: str) -> str:
          """Fill masked tokens in text using DeepInfra OpenAI-compatible API."""
          model = DEFAULT_MODELS["fill_mask"]
          prompt = f"""Fill in the [MASK] token in the following text with the most appropriate word. Provide the completed sentence and explain your choice.
  
  Text: {text}
  
  Response format: {{"filled_text": "completed sentence", "chosen_word": "word", "explanation": "reasoning"}}"""
          try:
              response = await client.completions.create(
                  model=model,
                  prompt=prompt,
                  max_tokens=200,
                  temperature=0.1,
              )
              if response.choices:
                  return response.choices[0].text
              else:
                  return "Unable to fill mask"
          except Exception as e:
              return f"Error filling mask: {type(e).__name__}: {str(e)}"

• src/mcp_deepinfra/server.py:264-265 (registration)

  Conditional block that registers the fill_mask tool with the FastMCP server using the @app.tool() decorator if enabled via ENABLED_TOOLS.

  if "all" in ENABLED_TOOLS or "fill_mask" in ENABLED_TOOLS:
      @app.tool()

• src/mcp_deepinfra/server.py:41-41 (helper)

  Configuration for the default model used by the fill_mask tool.

  "fill_mask": os.getenv("MODEL_FILL_MASK", "microsoft/DialoGPT-medium"),