Skip to content

adding-a-new-tool.md

Latest commit

 

History

History
89 lines (61 loc) · 3.06 KB

adding-a-new-tool.md

File metadata and controls

89 lines (61 loc) · 3.06 KB
Raw

Adding a New Tool to the MCP Server

To add a new capability to the Earthdata MCP server, follow these steps to build and register a new tool.

1. Create the Tool Directory

Create a new folder under tools/ with your tool's name (e.g., tools/get_weather/).

2. Add Required Files

You must include at least three files in your new directory:

A. manifest.json This file defines the tool's metadata for the MCP client. The loader automatically registers the tool using this file.

{
    "name": "get_weather",
    "description": "A detailed description of when the LLM should use this tool.",
    "version": "0.1.0",
    "entry_function": "get_weather"
}
  • Important: You MUST include a "version" string following Semantic Versioning (SemVer).

B. tool.py The actual Python implementation.

  • The primary function name must exactly match the "entry_function" in your manifest.json.
  • The function MUST be synchronous (def, not async def). The server automatically wraps it in an async handler.
from models.tools.get_weather import GetWeatherInput, GetWeatherOutput

def get_weather(location: str) -> dict:
    """Fetches the weather for a location."""
    # FastMCP infers the input schema from the function signature.
    # Using Pydantic parameters (or GetWeatherInput) is recommended.
    return GetWeatherOutput(status="success", temperature=72).model_dump()

C. output_model.py A Pydantic output model. Keep exactly one BaseModel subclass in this file; the loader registers the first matching model it finds when inspecting the module.

from pydantic import BaseModel

class GetWeatherOutput(BaseModel):
    status: str
    temperature: int

3. Update the Dockerfile & Dockerignore (Crucial Step)

If you added a new root-level Python directory (e.g., you created a prompts/ or config/ folder outside of tools/), you MUST update both the Dockerfile and the .dockerignore file.

If you forget either of these, the AWS ECS container will crash with a ModuleNotFoundError on startup because the strict ignore policy will strip the files from the build context!

1. Add to McpServerDockerfile:

COPY models/ models/
COPY tools/ tools/
COPY util/ util/
COPY new_folder/ new_folder/ # Add your new folder here!

2. Allowlist in McpServerDockerfile.dockerignore:

# util/
!util/

# new_folder/
!new_folder/ # Add your new folder here!

4. Tool Versioning

Every tool must include a "version" string in its manifest.json following Semantic Versioning (SemVer). You must bump this version whenever you change the tool's inputs, outputs, or LLM-facing behavior.

See Versioning Methodology for detailed policies on when to bump Major, Minor, or Patch versions for individual tools vs. the entire MCP Server.

5. Testing

  1. Run the server locally (uv run server.py http).
  2. Launch the MCP inspector (npx @modelcontextprotocol/inspector) to verify the LLM can see your tool's schema.
  3. Write pytest coverage under the tests/ directory.