GitHub MCP server with a Python agent

GitHub hosts an MCP server, so an agent can read repositories without custom API code. These two community-contributed scripts point a 10xGraph ReactAgent at that server and ask it questions about a repository. The agent uses OpenAI gpt-5-mini. For the basics of MCP clients, see the MCP overview.

What you will build

Two small scripts that point a ReactAgent at GitHub’s hosted MCP server. git_mcp.py asks for the commits of a repository and mcp_file_download.py reads its README.md.

Key concepts

fastmcp.Client with a remote streamable-http server and an Authorization header.
ReactAgent(client=...) to expose the server’s tools to the model.
InMemoryCheckpointer to keep the thread state between steps.

Run it

Terminal

git clone https://github.com/10xGraph/10xGraph.git
cd 10xGraph/agentflow
pip install "10xgraph[mcp,openai]" python-dotenv
export OPENAI_API_KEY=...
export GITHUB_TOKEN=...   # a personal access token
python examples/mcp/github/git_mcp.py
python examples/mcp/github/mcp_file_download.py

How the code works

`git_mcp.py`

Python

mcp_config = {
    "mcpServers": {
        "github": {
            "url": "https://api.githubcopilot.com/mcp/",
            "headers": {"Authorization": f"Bearer {os.environ['GITHUB_TOKEN']}"},
            "transport": "streamable-http",
        },
    }
}

The token goes in an Authorization header, so it never appears in the prompt or the message history.

Python

agent = ReactAgent(
    model="gpt-5-mini",
    system_prompt=[
        {
            "role": "system",
            "content": "You are a helpful assistant. Use the GitHub tools to answer questions.",
        }
    ],
    client=Client(mcp_config),  # tools are discovered from the server at run time
    trim_context=True,
)

No tool is written by hand. trim_context=True keeps long tool results, like a full commit list, from filling the context window.

Python

    app = agent.compile(checkpointer=InMemoryCheckpointer())
    question = (
        "Call the list_commits tool for the repo 'portfolio' owned by 'suchith83' "
        "and give me all commits in that repo."
    )
    result = await app.ainvoke(
        {"messages": [Message.text_message(question)]},
        config={"thread_id": "github-commits", "recursion_limit": 10},
    )
    for message in result["messages"]:
        print(f"[{message.role}]", message.text()[:500])

The model decides to call list_commits, ReactAgent runs it through the MCP client, and the loop ends when the model answers in text. The script prints every message, cut to 500 characters, so you can see the tool result.

`mcp_file_download.py`

It builds the same agent. Only the question and thread id change.

Python

    question = "Get the README.md file from the repo 'portfolio' owned by 'suchith83'."
    result = await app.ainvoke(
        {"messages": [Message.text_message(question)]},
        config={"thread_id": "github-readme", "recursion_limit": 10},
    )
    print("assistant:", result["messages"][-1].text())

The agent picks the right GitHub tool on its own. It prints only the final answer.

Full code

git_mcp.py

Python

"""List commits from a GitHub repository through GitHub's remote MCP server.

A ReactAgent gets its tools from the MCP server, so no tool is written by hand.
Contributed example, modernised to use ReactAgent(client=...).

Run: GITHUB_TOKEN=... OPENAI_API_KEY=... python examples/mcp/github/git_mcp.py
"""

import asyncio
import os

from dotenv import load_dotenv
from fastmcp import Client

from tenxgraph.core.state import Message
from tenxgraph.prebuilt.agent import ReactAgent
from tenxgraph.storage.checkpointer import InMemoryCheckpointer


load_dotenv()

# GitHub hosts an MCP server over streamable HTTP. The token goes in an
# Authorization header, so it never appears in the prompt.
mcp_config = {
    "mcpServers": {
        "github": {
            "url": "https://api.githubcopilot.com/mcp/",
            "headers": {"Authorization": f"Bearer {os.environ['GITHUB_TOKEN']}"},
            "transport": "streamable-http",
        },
    }
}

agent = ReactAgent(
    model="gpt-5-mini",
    system_prompt=[
        {
            "role": "system",
            "content": "You are a helpful assistant. Use the GitHub tools to answer questions.",
        }
    ],
    client=Client(mcp_config),  # tools are discovered from the server at run time
    trim_context=True,
)


async def main() -> None:
    app = agent.compile(checkpointer=InMemoryCheckpointer())
    question = (
        "Call the list_commits tool for the repo 'portfolio' owned by 'suchith83' "
        "and give me all commits in that repo."
    )
    result = await app.ainvoke(
        {"messages": [Message.text_message(question)]},
        config={"thread_id": "github-commits", "recursion_limit": 10},
    )
    for message in result["messages"]:
        print(f"[{message.role}]", message.text()[:500])
    await app.aclose()


if __name__ == "__main__":
    asyncio.run(main())

mcp_file_download.py

Python

"""Fetch a file (a repository README) through GitHub's remote MCP server.

Same setup as git_mcp.py with a different question: the agent picks the right
MCP tool on its own.

Run: GITHUB_TOKEN=... OPENAI_API_KEY=... python examples/mcp/github/mcp_file_download.py
"""

import asyncio
import os

from dotenv import load_dotenv
from fastmcp import Client

from tenxgraph.core.state import Message
from tenxgraph.prebuilt.agent import ReactAgent
from tenxgraph.storage.checkpointer import InMemoryCheckpointer


load_dotenv()

mcp_config = {
    "mcpServers": {
        "github": {
            "url": "https://api.githubcopilot.com/mcp/",
            "headers": {"Authorization": f"Bearer {os.environ['GITHUB_TOKEN']}"},
            "transport": "streamable-http",
        },
    }
}

agent = ReactAgent(
    model="gpt-5-mini",
    system_prompt=[
        {
            "role": "system",
            "content": "You are a helpful assistant. Use the GitHub tools to answer questions.",
        }
    ],
    client=Client(mcp_config),
    trim_context=True,
)


async def main() -> None:
    app = agent.compile(checkpointer=InMemoryCheckpointer())
    question = "Get the README.md file from the repo 'portfolio' owned by 'suchith83'."
    result = await app.ainvoke(
        {"messages": [Message.text_message(question)]},
        config={"thread_id": "github-readme", "recursion_limit": 10},
    )
    print("assistant:", result["messages"][-1].text())
    await app.aclose()


if __name__ == "__main__":
    asyncio.run(main())

Source on GitHub: examples/mcp/github

Expected output

Illustrative:

Text

[user] Call the list_commits tool for the repo 'portfolio' ...
[assistant]
[tool] [{"sha": "3f2a...", "commit": {"message": "Add contact form"}}, ...]
[assistant] The repo has 12 commits. The latest is "Add contact form".

Using a different model

Change gpt-5-mini to gemini-2.5-flash (pip install "10xgraph[mcp,google-genai]") or claude-sonnet-5-5 (pip install "10xgraph[mcp,anthropic]") and set that provider’s key.

Next steps

MCP overview
Xquik MCP for an API key header instead of a bearer token.
Providers for other models.

Frequently asked questions

Where does the GitHub token go?

In the headers of the server entry in the mcpServers config, as an Authorization bearer token read from GITHUB_TOKEN. It is never part of the prompt.

Do I write the GitHub tools myself?

No. ReactAgent(client=...) discovers the tools the GitHub MCP server exposes, such as list_commits, at run time.