MCP Integration

RoomKit is an MCP client: its AI channels call tools that Model Context Protocol servers expose. RoomKit does not ship an MCP server of its own.

MCP tools in an AI channel

pip install "roomkit[mcp]"

MCPToolProvider connects to an MCP server, by URL or by starting it as a command, discovers its tools, and hands them to an AIChannel as ordinary RoomKit tools. compose_tool_handlers mixes them with your local tools:

from roomkit import AIChannel
from roomkit.tools import MCPToolProvider, compose_tool_handlers

async with MCPToolProvider.from_url("http://localhost:8000/mcp") as mcp:
    handler = compose_tool_handlers(local_handler, mcp.as_tool_handler())
    ai = AIChannel("ai-assistant", provider=provider, tool_handler=handler)
    # register the channel, then attach it with mcp.get_tools_as_dicts() in its metadata

An MCP tool call passes the same gates as a local one: tool policies and ON_TOOL_CALL hooks. The MCP Tool Provider guide covers transports, filtering, headers and the tool-name aliases.

MCP servers for an ACP agent

An external agent that takes part in a room over ACP (Claude Code, for example) brings its own MCP servers. ACPChannel(mcp_servers=[...]) declares them in the agent's session; they name servers on the agent's machine, not on RoomKit's. See the ACP Agent Channel guide.

Driving RoomKit from an MCP client

To let an MCP client such as Claude Desktop act on rooms, write a server with the MCP Python SDK whose tools call RoomKit's public API: kit.create_room(), kit.get_room(), kit.attach_channel(), kit.get_timeline(), kit.process_inbound(). Check who the caller is before each call: the server acts with the rights you give it.