Give AI Access to an API

Wrap a real REST API as MCP tools so the assistant never invents URLs or auth headers.

The model should not concatenate https://api.github.com/ and hope. You wrap the endpoints you trust, with timeouts and a User-Agent, and you return a thin result.

Pick an API you can call without a secret first

Use a public JSON endpoint so this lesson is not blocked on keys. https://httpbin.org/get echoes the request. GitHub’s public user API works without a token for a single profile. We will use httpbin so you are not fighting rate limits.

python -m pip install httpx

One tool, one HTTP call

import httpx
from mcp.server.mcpserver import MCPServer

mcp = MCPServer("http-demo")


@mcp.tool()
def inspect_get(url: str = "https://httpbin.org/get") -> dict:
    '''GET a URL and return status plus JSON or a text preview. Default is httpbin.'''
    headers = {"User-Agent": "geek-university-mcp-course/1.0"}
    with httpx.Client(timeout=10.0, headers=headers, follow_redirects=True) as client:
        response = client.get(url)
    payload: dict = {"status": response.status_code, "url": str(response.url)}
    try:
        payload["json"] = response.json()
    except ValueError:
        payload["text"] = response.text[:500]
    return payload


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

Do not take a free-form url in a production tool. The default above is for class. In a real server, hard-code the host or allow a small allow-list. Open URL tools become SSRF.

What “wrap” means

The tool name is a verb the model understands (inspect_get), not get_endpoint.
You own retries, timeouts, and which fields come back. Drop headers and cookies the model does not need.
Auth, when you add it, lives in the environment — see Authentication and Secrets.

Check it

Invoke the tool with the default URL. You should see status 200 and a JSON body that includes a url field. Then lock the URL to a constant and remove the argument — that is the shape of a useful wrapper. Next: validation and error shapes.