MCP Overview

The Model Context Protocol lets a chain call tools hosted by an external MCP server. MCPStepDescription invokes one such tool. (In v0.3.0 MCP graduated from experimental.)

pip install 'mmar-carl[mcp]'

from mmar_carl import MCPStepDescription, MCPStepConfig, MCPServerConfig

MCPStepDescription(
    number=1,
    title="Search via MCP",
    config=MCPStepConfig(
        server=MCPServerConfig(
            server_name="my_server",
            command="python",
            args=["-m", "my_mcp_server"],   # stdio transport
        ),
        tool_name="search",
        argument_mapping={"query": "$history[-1]"},
    ),
)

MCPServerConfig

Field	Default	Purpose
server_name	—	Logical name for the server.
transport	"stdio"	"stdio" (subprocess), "http" (streamable HTTP), "sse" (HTTP + Server-Sent Events).
command / args	—	How to launch a stdio server.
url	—	Base URL for http / sse.
headers	{}	Extra HTTP headers (e.g. Authorization) for http / sse.

MCPStepConfig

Field	Default	Purpose
server	—	The MCPServerConfig.
tool_name	—	MCP tool to call.
arguments	{}	Static arguments.
argument_mapping	{}	Arguments resolved from references.
timeout	60.0	Seconds.

Connections are managed by an internal pool (mcp_pool), and MCP runs support pause/resume, cancellation, streaming, and lossless serialization.

Tip

For production, prefer registering MCP capabilities as ordinary tools via context.register_tool(...) when you can — fewer moving parts than a live server connection.

See also

• MCP resources — read data (not call tools) from a server.