ChatGPT 🤝 FastMCP

ChatGPT supports MCP servers through remote HTTP connections, allowing you to extend ChatGPT’s capabilities with custom tools and knowledge from your FastMCP servers.

​

Deep Research

ChatGPT’s Deep Research feature requires MCP servers to be internet-accessible HTTP endpoints with exactly two specific tools:

• search: For searching through your resources and returning matching IDs
• fetch: For retrieving the full content of specific resources by ID

​

Tool Descriptions Matter

Since ChatGPT needs to understand how to use your tools effectively, write detailed tool descriptions. The description teaches ChatGPT how to form queries, what parameters to use, and what to expect from your data. Poor descriptions lead to poor search results.

​

Create a Server

A Deep Research-compatible server must implement these two required tools:

• search(query: str) - Takes a query of any kind and returns matching record IDs
• fetch(id: str) - Takes an ID and returns the record

Critical: Write detailed docstrings for both tools. These descriptions teach ChatGPT how to use your tools effectively. Poor descriptions lead to poor search results.

The search tool should take a query (of any kind!) and return IDs. The fetch tool should take an ID and return the record.

Here’s a reference server implementation you can adapt (see also OpenAI’s sample server for comparison):

import json
from pathlib import Path
from dataclasses import dataclass
from fastmcp import FastMCP

@dataclass
class Record:
    id: str
    title: str
    text: str
    metadata: dict

def create_server(
    records_path: Path | str,
    name: str | None = None,
    instructions: str | None = None,
) -> FastMCP:
    """Create a FastMCP server that can search and fetch records from a JSON file."""
    records = json.loads(Path(records_path).read_text())

    RECORDS = [Record(**r) for r in records]
    LOOKUP = {r.id: r for r in RECORDS}

    mcp = FastMCP(name=name or "Deep Research MCP", instructions=instructions)

    @mcp.tool()
    async def search(query: str):
        """
        Simple unranked keyword search across title, text, and metadata.
        Searches for any of the query terms in the record content.
        Returns a list of matching record IDs for ChatGPT to fetch.
        """
        toks = query.lower().split()
        ids = []
        for r in RECORDS:
            record_txt = " ".join(
                [r.title, r.text, " ".join(r.metadata.values())]
            ).lower()
            if any(t in record_txt for t in toks):
                ids.append(r.id)

        return {"ids": ids}

    @mcp.tool()
    async def fetch(id: str):
        """
        Fetch a record by ID.
        Returns the complete record data for ChatGPT to analyze and cite.
        """
        if id not in LOOKUP:
            raise ValueError(f"Unknown record ID: {id}")
        return LOOKUP[id]

    return mcp

if __name__ == "__main__":
    mcp = create_server("path/to/records.json")
    mcp.run(transport="http", port=8000)

​

Deploy the Server

Your server must be deployed to a public URL in order for ChatGPT to access it.

For development, you can use tools like ngrok to temporarily expose a locally-running server to the internet. We’ll do that for this example (you may need to install ngrok and create a free account), but you can use any other method to deploy your server.

Assuming you saved the above code as server.py, you can run the following two commands in two separate terminals to deploy your server and expose it to the internet:

python server.py

​

Connect to ChatGPT

Replace https://your-server-url.com with the actual URL of your server (such as your ngrok URL).

1. Open ChatGPT and go to Settings → Connectors
2. Click Add custom connector
3. Enter your server details:
   • Name: Library Catalog
   • URL: Your server URL, including the path.
     • Note: Ensure your URL includes the correct path for the transport you’re using. The defaults are /sse/ for SSE (e.g., https://abc123.ngrok.io/sse/) and /mcp/ for HTTP (e.g., https://abc123.ngrok.io/mcp/).
   • Description: A library catalog for searching and retrieving books

​

Test the Connection

1. Start a new chat in ChatGPT
2. Click Tools → Run deep research
3. Select your Library Catalog connector as a source
4. Ask questions like:
   • “Search for Python programming books”
   • “Find books about AI and machine learning”
   • “Show me books by the Python Software Foundation”

ChatGPT will use your server’s search and fetch tools to find relevant information and cite the sources in its response.

​

Troubleshooting

​

”This MCP server doesn’t implement our specification”

If you get this error, it most likely means that your server doesn’t implement the required tools (search and fetch). To correct it, ensure that your server meets the service requirements.