Science MCP Server Starter

Minimal starter code for wrapping science functions as MCP tools.

The science code stays normal Python. The MCP wrapper only imports the public tool functions listed in pyproject.toml and exposes them through FastMCP.

This repo only defines the MCP server. Agent code lives in the client repo that connects to it.

Connect Science Code

1. Copy the MCP wrapper into the science repo root.

mcp_server/

This repo uses science_demo/ as the example science package.

2. Create normal Python functions in the science package.

def plot_sine_wave(output_dir: str, num_points: int = 200):
    """Plot a sine wave and save it in output_dir."""
    ...

3. Export only the MCP tools from the science package __init__.py.

from .science_tools import plot_sine_wave

__all__ = ["plot_sine_wave"]

Only names in __all__ become MCP tools. Helper functions stay private. This starter does not scan a repo for functions.

4. Add the MCP dependency and tool module config in pyproject.toml.

dependencies = [
    "mcp[cli]>=1.27,<2",
]

[tool.mcp-server]
tool_modules = ["your_science_package"]

5. Install the science repo.

python3.12 -m venv .venv
source .venv/bin/activate
pip install -e .

6. Connect an agent with one transport.

Use stdio for a local MCP server that the agent's MCP client starts for you. Do not start the stdio server in a separate terminal. Configure the client with the command and arguments to launch:

{
  "science": {
    "transport": "stdio",
    "command": "python",
    "args": ["-m", "mcp_server", "--transport", "stdio"]
  }
}

Use streamable-http when the agent connects to an already running server. Start the server in a terminal:

python -m mcp_server --transport streamable-http --host 127.0.0.1 --port 8000

Then point the agent at:

http://127.0.0.1:8000/mcp

Related MCP server: openclaw-tools-mcp

Demo Tools

The demo science package exposes:

generate_random_points(output_dir: str, count: int = 30, seed: int = 42) -> ArtifactResult
plot_sine_wave(output_dir: str, num_points: int = 200) -> ArtifactResult

The type hints, docstrings, and Pydantic constraints become the MCP tool schema that external agents see.

Successful artifact tools return:

{"status": "success", "files": ["..."], "message": "...", "metadata": {}}