Skip to content

agentix.mcp

mcp

MCP (Model Context Protocol) client support.

Connect to an MCP server, discover its tools, and expose them as agentix :class:~agentix.tools.Tool objects that route calls back over the live connection. The discovered tools plug straight into Agent(tools=...)::

async with MCPServer(command="my-mcp-server", args=["--flag"]) as server:
    agent = Agent(model=..., system_prompt="...", tools=await server.list_tools())
    await agent.run("...")            # run INSIDE the block

The tools are only valid while the connection is open, so keep the agent run inside the async with. Requires the mcp package (lazy-imported): pip install "agentix[mcp]".

MCP tools already carry a JSON Schema (inputSchema), which is exactly agentix's parameters convention — so they flow through to any provider adapter (e.g. Anthropic's input_schema) unchanged.

MCPServer 

MCPServer(
    *,
    command: str | None = None,
    args: Sequence[str] | None = None,
    env: dict[str, str] | None = None,
    url: str | None = None,
    transport: str | None = None,
    name: str | None = None,
    session: Any = None,
)

A connection to a single MCP server.

Construct with a transport (command= for stdio, url= for HTTP/SSE), or pass a pre-built session (anything exposing async initialize() / list_tools() / call_tool(name, args)) for testing or to reuse an existing client. Use as an async context manager.

list_tools async 

list_tools() -> list[Tool]

Discover the server's tools as agentix :class:Tool objects.