.. Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements. See the NOTICE file distributed with this work for additional information regarding copyright ownership. The ASF licenses this file to you under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

.. http://www.apache.org/licenses/LICENSE-2.0

.. Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.

.. _howto/connection:mcp:

MCP Server Connection

The MCP connection type configures access to MCP (Model Context Protocol) <https://modelcontextprotocol.io/>__ servers. Three transport types are supported: Streamable HTTP, SSE, and stdio.

Default Connection IDs

The MCPHook uses mcp_default by default.

Configuring the Connection

Transport (Extra field) The transport type: http (default), sse, or stdio.

- ``http``: Streamable HTTP — the recommended transport for remote servers.
- ``sse``: Server-Sent Events — deprecated in favor of Streamable HTTP.
- ``stdio``: Run the MCP server as a subprocess communicating over stdin/stdout.

Host The server URL. Required for http and sse transports.

Examples: ``http://localhost:3001/mcp``, ``https://mcp.example.com/v1``

Auth Token (Password field) Optional authentication token for the MCP server. Sent as a static Authorization: Bearer <token> header on HTTP/SSE transports. For short-lived or minted tokens, use a token_provider instead (see below).

Command (Extra field) The command to run for stdio transport. Required when transport is stdio.

Examples: ``uvx``, ``python``, ``node``

Arguments (Extra field) JSON array of arguments for the stdio command.

Examples: ``["mcp-run-python"]``, ``["-m", "my_mcp_server"]``

Examples

HTTP transport (remote MCP server)

.. code-block:: json

{
    "conn_type": "mcp",
    "host": "http://localhost:3001/mcp"
}

SSE transport

.. code-block:: json

{
    "conn_type": "mcp",
    "host": "http://localhost:3001/sse",
    "extra": "{\"transport\": \"sse\"}"
}

Stdio transport (subprocess)

.. code-block:: json

{
    "conn_type": "mcp",
    "extra": "{\"transport\": \"stdio\", \"command\": \"uvx\", \"args\": [\"mcp-run-python\"]}"
}

Stdio with custom timeout

.. code-block:: json

{
    "conn_type": "mcp",
    "extra": "{\"transport\": \"stdio\", \"command\": \"python\", \"args\": [\"-m\", \"my_server\"], \"timeout\": 30}"
}

Short-lived or minted tokens

Some MCP endpoints require a freshly minted, short-lived token rather than a static one. For example, Snowflake managed MCP servers <https://docs.snowflake.com/en/user-guide/snowflake-cortex/cortex-agents-mcp>__ are best authenticated with a key-pair JWT <https://docs.snowflake.com/en/user-guide/key-pair-auth>__: the private key never leaves your environment and the signed JWT expires after about an hour, so it cannot be stored as a static connection password. The same applies to OAuth / refresh tokens, Workload Identity Federation, and GitHub App installation tokens.

For these, pass a token_provider callable to MCPHook or MCPToolset instead of a static token. It is called each time the connection is established and its return value is used as the bearer token, so a fresh token is minted (and registered with secret masking so it does not leak into task logs):

.. code-block:: python

from airflow.providers.common.ai.toolsets.mcp import MCPToolset


def mint_snowflake_jwt() -> str:
    # Sign a short-lived JWT from the Snowflake connection's key-pair.
    ...


toolset = MCPToolset(
    mcp_conn_id="snowflake_managed_mcp",
    token_provider=mint_snowflake_jwt,
)

token_provider is resolved in DAG code (it is a Python callable, not a stored connection field), so the signing key stays in your environment and is never baked into the serialized DAG.