Skip to main content

Agent connector tutorial: FastMCP

In this tutorial, you'll create a new Python project with uv, build a FastMCP server that exposes one of Airbyte's agent connectors as an MCP tool, and use it to query GitHub data from any MCP-compatible agent. This tutorial uses GitHub, but if you don't have a GitHub account you can swap in any other agent connector and perform different operations.

Your MCP server executes through Airbyte. Airbyte Agents owns the OAuth apps, stores your third-party tokens, and refreshes them for you. Your Python code only ever sees your Airbyte client ID and client secret.

Overview​

This tutorial is for AI engineers and other technical users who work with data and AI tools. You can complete it in about 15 minutes.

The tutorial assumes you have basic knowledge of the following tools, but most software engineers shouldn't struggle with anything that follows.

  • Python and package management with uv
  • MCP (Model Context Protocol) and MCP servers
  • GitHub, or a different third-party service you want to connect to

Before you start​

Before you begin this tutorial, ensure you have the following.

Part 1: Create a new Python project​

Create a new project using uv:

uv init my-mcp-agent --app
cd my-mcp-agent

This creates a project with the following structure:

my-mcp-agent/
├── .gitignore
├── .python-version
├── main.py
├── pyproject.toml
└── README.md

Part 2: Install dependencies​

Install the Airbyte agent SDK and FastMCP:

uv add airbyte-agent-sdk fastmcp python-dotenv

This command installs:

  • airbyte-agent-sdk: The Airbyte Agents Python SDK, which ships every connector as a typed submodule.
  • fastmcp: A Python framework for building MCP servers with minimal boilerplate.
  • python-dotenv: A library you can use to load environment variables from a .env file.

Part 3: Import FastMCP and the GitHub agent connector​

  1. Create a server.py file for your MCP server definition:

    touch server.py

  2. Add the following imports to server.py:

    server.py
    import json

    from dotenv import load_dotenv
    from fastmcp import FastMCP
    from airbyte_agent_sdk import connect
    from airbyte_agent_sdk.connectors.github import GithubConnector

    These imports provide:

    • json: Serialize connector results for the MCP tool return value.
    • load_dotenv: Load environment variables from your .env file.
    • FastMCP: The FastMCP server class that handles MCP protocol communication.
    • connect: The Airbyte agent SDK entry point. One call returns a typed connector bound to your workspace.
    • GithubConnector: The connector class. You reference it when decorating the tool so the SDK can describe the connector's entities and actions to the agent.

Part 4: Add a .env file with your secrets​

  1. Create a .env file in your project root and add your Airbyte API credentials to it. Replace the placeholder values with your actual credentials.

    .env
    AIRBYTE_CLIENT_ID=your-airbyte-client-id
    AIRBYTE_CLIENT_SECRET=your-airbyte-client-secret

    Copy these values from the Profile page in the Airbyte Agents web app.

    warning

    Never commit your .env file to version control. If you do this by mistake, rotate your secrets immediately.

  2. Add the following line to server.py after your imports to load the environment variables:

    server.py
    load_dotenv()

Part 5: Configure your connector and MCP server​

Now that your environment is set up, add the following code to server.py to create the GitHub connector and FastMCP server.

Create the server and connector​

server.py
mcp = FastMCP("GitHub Agent")

github = connect("github")

FastMCP("GitHub Agent") creates a new MCP server named "GitHub Agent".

connect("github") does four things for you:

  • Reads AIRBYTE_CLIENT_ID and AIRBYTE_CLIENT_SECRET from the environment.
  • Defaults to the "default" workspace, which is where the web app stores credentials unless you change it.
  • Returns a typed GithubConnector bound to the authenticated GitHub connector you added earlier.
  • Routes every github.execute(...) call through Airbyte's hosted API, which holds the GitHub OAuth tokens and refreshes them for you.

You never register an OAuth app, copy a GitHub token into your code, or write token-refresh logic.

If you want to connect to a different workspace or pass credentials explicitly, use connect("github", workspace_name="my-workspace", client_id=..., client_secret=...) or pass an AirbyteAuthConfig. See the SDK reference for details.

Register the tool​

Rather than one tool per GitHub endpoint, the Airbyte agent SDK exposes the entire GitHub API through a single execute(entity, action, params) entry point. The @GithubConnector.tool_utils decorator fills in the entity and action catalog as part of the tool description, so the agent knows what's available without you writing a schema.

server.py
@mcp.tool()
@GithubConnector.tool_utils
async def github_execute(entity: str, action: str, params: dict | None = None) -> str:
    """Execute GitHub connector operations."""
    result = await github.execute(entity, action, params or {})
    return json.dumps(result, default=str)

The decorator stack is the whole tool definition. No per-action docstring, no GITHUB_LIST_COMMITS or GITHUB_GET_PR sprawl, one entry point that covers the full connector. @GithubConnector.tool_utils appends the full entity and action catalog to the tool description so the MCP client sees every entity, action, and enum value the connector supports. As the connector grows, the tool signature stays the same.

Each execute call returns a structured result with data (the records) and meta (pagination cursors). MCP tools return strings, so this tutorial serializes the whole result with json.dumps so the MCP client can reason about both the records and the pagination state.

Add the server entry point​

Add the following at the bottom of server.py to start the server when run directly:

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

Part 6: Register with your agent​

Register the MCP server with your preferred agent. Provide the full path to your project's server.py file. Replace /path/to/my-mcp-agent with the actual path to your project directory.

claude mcp add github-agent -- uv run --directory /path/to/my-mcp-agent server.py

Part 7: Use the MCP server​

  1. Restart your agent so it picks up the new MCP server registration.

  2. Once restarted, prompt your agent with natural language questions about your GitHub data. Try prompts like:

    • "List the 10 most recent open issues in airbytehq/airbyte"
    • "What are the 10 most recent pull requests that are still open in airbytehq/airbyte?"
    • "Are there any open issues that might be fixed by a pending PR?"

Your agent discovers the MCP server's tools automatically and calls them based on your prompts. The MCP server hands each tool call off to Airbyte, which executes the operation against GitHub and returns the result.

Troubleshooting​

If your agent fails to retrieve GitHub data, check the following:

  • Server not found: Ensure the path in your MCP configuration points to the correct server.py file and that uv is available on your system PATH.
  • HTTP 401/403 errors from Airbyte: Verify that AIRBYTE_CLIENT_ID and AIRBYTE_CLIENT_SECRET are copied correctly from your Profile page.
  • "No connector found" or "connector not configured": Make sure you've added a GitHub connector in the Credentials page of the Airbyte Agents web app. connect("github") defaults to the "default" workspace; if you added the connector to a different workspace, pass workspace_name="your-workspace-name" to connect().
  • HTTP 401/403 errors from GitHub: The GitHub token or OAuth credentials stored in your connector are invalid or missing required scopes. Open your GitHub connector in the web app and reauthenticate with a valid token that has repo scope.
  • Empty data=[] responses from filtered queries: Most GitHub filters use case-sensitive values. Confirm the agent is sending uppercase values (for example, states=["OPEN"] rather than states=["open"]). The tool description's rules nudge the model to do that by default; you can also reinforce the rules in your client's system prompt.

See the Github agent connector page for more details.

Summary​

In this tutorial, you learned how to:

  • Set up a new Python project with uv
  • Add FastMCP and Airbyte's GitHub agent connector to your project
  • Configure environment variables for your Airbyte Agents credentials
  • Register a single MCP tool that covers the entire GitHub API
  • Register your MCP server with an agent and use natural language to interact with GitHub data through Airbyte

Next steps​

  • Add another connector. The same connect(...) + execute(...) pattern covers the full Airbyte agent connectors catalog. Add Slack, Stripe, Salesforce, or any other connector in the web app, then call slack = connect("slack") in your server and register a second tool with another @mcp.tool() / @SlackConnector.tool_utils stack. Your MCP client now reads GitHub and posts to Slack with no additional OAuth setup.
  • Use write actions. Connectors expose create, update, and post actions alongside the read ones. Ask your client to file an issue, comment on a PR, or send a Slack message, and execute carries the write through with the stored OAuth token.
  • Let your AI assistant scaffold the next server. The Airbyte agent SDK ships skills for Claude Code and Codex that carry the patterns above, so you can ask your assistant to build a new MCP server without retyping them. See the airbyte-agent-sdk repository for installation instructions.
  • Reach the same connectors from a hosted MCP endpoint. Airbyte Agents exposes the same connectors through a hosted MCP endpoint that works with Claude Code, Cursor, and ChatGPT, with one OAuth flow per provider shared across clients. Use this when you don't want to run and maintain your own MCP server.
Was this page helpful?