Ship an MCP Server in Python That Exposes Your Internal API to LLMs

What You'll Build

A Python MCP server using FastMCP that wraps a corporate REST API as three structured tools—search_customers, get_order, and create_support_ticket. Any MCP-compatible client (Claude Desktop, Cursor, custom agents) can call your API with full type safety, without the model ever seeing credentials or constructing raw URLs.

Prerequisites

• Python 3.10+ (required for built-in generic types like list[dict])
• pip or uv for package management
• Node.js 18+ — mcp dev invokes npx @modelcontextprotocol/inspector under the hood
• Latest Claude Desktop (for end-to-end testing; optional if using only the inspector)
• A REST API with a bearer token — a mock URL works fine to follow along
• Comfortable with async/await Python

1. Set Up the Project

mkdir mcp-internal-api && cd mcp-internal-api
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\activate
pip install "mcp[cli]" httpx python-dotenv

mcp[cli] installs the mcp CLI used for local inspection. httpx handles async HTTP to your backend.

Create .env for local credentials — add it to .gitignore now:

API_BASE_URL=https://api.corp.example.com
API_KEY=sk-your-real-token-here

2. Write the Server

Create server.py:

import os
import httpx
from dotenv import load_dotenv
from mcp.server.fastmcp import FastMCP

load_dotenv()

mcp = FastMCP("internal-api")

_BASE = os.environ["API_BASE_URL"]
_KEY  = os.environ["API_KEY"]


def _auth_headers() -> dict[str, str]:
    return {"Authorization": f"Bearer {_KEY}", "Accept": "application/json"}


@mcp.tool()
async def search_customers(query: str, limit: int = 10) -> list[dict]:
    """Search customers by name or email. Returns a list of customer records."""
    async with httpx.AsyncClient() as client:
        r = await client.get(
            f"{_BASE}/customers",
            headers=_auth_headers(),
            params={"q": query, "limit": limit},
            timeout=10.0,
        )
        r.raise_for_status()
        return r.json()


@mcp.tool()
async def get_order(order_id: str) -> dict:
    """Fetch a single order by its ID."""
    async with httpx.AsyncClient() as client:
        r = await client.get(
            f"{_BASE}/orders/{order_id}",
            headers=_auth_headers(),
            timeout=10.0,
        )
        r.raise_for_status()
        return r.json()


@mcp.tool()
async def create_support_ticket(
    customer_id: str,
    subject: str,
    body: str,
    priority: str = "normal",
) -> dict:
    """Open a support ticket for a customer.

    Args:
        customer_id: The customer's UUID.
        subject: One-line summary (max 120 chars).
        body: Full description of the issue.
        priority: 'low', 'normal', or 'high'.
    """
    if priority not in {"low", "normal", "high"}:
        raise ValueError(f"priority must be low/normal/high, got '{priority}'")

    async with httpx.AsyncClient() as client:
        r = await client.post(
            f"{_BASE}/tickets",
            headers=_auth_headers(),
            json={
                "customer_id": customer_id,
                "subject": subject,
                "body": body,
                "priority": priority,
            },
            timeout=10.0,
        )
        r.raise_for_status()
        return r.json()


if __name__ == "__main__":
    mcp.run()

Why each decision matters:

mcp.run() defaults to stdio transport, which is what Claude Desktop and most local clients expect — the client spawns your server as a subprocess and talks JSON-RPC over stdin/stdout.

3. Inspect Locally with mcp dev

Before touching any LLM, validate the wiring in a browser UI:

mcp dev server.py

This starts your server and opens the MCP Inspector (the URL is printed in your terminal). Navigate to Tools — you'll see all three tools with auto-generated input forms matching your Python signatures. Call search_customers with query = "alice" and confirm a JSON response or a typed upstream error.

Tip: Set API_BASE_URL=https://httpbin.org temporarily to exercise the async/auth plumbing without a live internal API. You'll get a 404 back, which correctly surfaces as an httpx.HTTPStatusError tool error.

4. Wire to Claude Desktop

Locate the config file:

Add your server entry. Use absolute paths — Claude Desktop spawns a clean, non-login shell that won't activate your virtualenv:

{
  "mcpServers": {
    "internal-api": {
      "command": "/absolute/path/to/.venv/bin/python",
      "args": ["/absolute/path/to/server.py"],
      "env": {
        "API_BASE_URL": "https://api.corp.example.com",
        "API_KEY": "sk-your-real-token-here"
      }
    }
  }
}

Restart Claude Desktop, then in a new conversation:

"Search for customers named 'smith', then open a high-priority support ticket for the first result explaining their order is delayed."

Claude will call search_customers, inspect the output, then call create_support_ticket — tool calls appear inline in the UI with their arguments and responses visible.

Verify It Works

Inspector: After mcp dev server.py, the Tools tab lists all three tools with correct schemas and no import errors in the terminal.

Claude Desktop: Open Settings → Developer. Your server appears as internal-api with a green connected indicator. If it shows an error state, restart Claude Desktop after editing the config.

Schema sanity-check — confirm FastMCP generated correct schemas without starting a full client:

python -c "
import os; os.environ['API_BASE_URL']='http://x'; os.environ['API_KEY']='x'
import asyncio
import server
async def main():
    for tool in await server.mcp.list_tools():
        print(tool.name, tool.inputSchema)
asyncio.run(main())
"

You should see each tool name alongside its JSON Schema inputSchema dict.

Troubleshooting

ModuleNotFoundError: No module named 'mcp' — Claude Desktop uses a clean shell; your virtualenv isn't activated. Confirm "command" points to the venv interpreter: /path/to/.venv/bin/python, not the system python.

Tools don't appear in Claude Desktop — Run mcp dev server.py first; import errors or missing env vars appear there immediately. Also check ~/Library/Logs/Claude/ on macOS — Claude Desktop writes a per-server log file named after your server key (internal-api).

KeyError: 'API_BASE_URL' — The env block in claude_desktop_config.json replaces the shell environment entirely; load_dotenv() won't read your .env from there. Set all required keys explicitly in the JSON config.

httpx.ReadTimeout — Your backend is slow. Raise timeout=30.0, or restructure long-running operations to return an AsyncGenerator and use yield to stream partial results back to the client.

Next Steps

• Resources: Expose read-only context (OpenAPI specs, internal wikis) via @mcp.resource() so the LLM can pull reference material without consuming tool-call budget.
• HTTP/SSE transport: For multi-user or remote deployments, replace mcp.run() with mcp.run(transport="sse") and mount it behind a secured reverse proxy; validate per-request tokens in middleware rather than a static env var.
• Rate limiting: Wrap _auth_headers() with a token-bucket limiter (aiolimiter is async-native) to prevent an agentic loop from flooding your upstream API.
• Richer schemas: Replace dict return types with Pydantic models — FastMCP generates detailed JSON Schema from them, giving the model better guidance on what fields to expect and use.
• Spec & SDK: modelcontextprotocol.io/docs and the Python SDK on GitHub.