cogno-mcp

Bridge an MCP server's tools to the Cogno tool contract.

The third tool source for the Cogno stack — alongside in-process skills (cogno-cortex) and the host's native functions. MCPDispatcher exposes any MCP server's tools as a cogno-anima ToolDispatcher, so the EGO / cogno-soma see MCP tools as ordinary tools and never speak the protocol.

The valuable part — schema conversion, CallToolResult → ToolResult mapping, and policy (read-only / destructive) from MCP tool annotations — is SDK-free and fully tested with fakes. The thin transport helpers (stdio / HTTP / SSE) are the only place the official mcp SDK is used; transport is the SDK's job, not ours.

MCP server ──(stdio/HTTP/SSE)──▶ MCPDispatcher ──▶ ToolDispatcher ──▶ EGO / cogno-soma
                                  (implements cogno-anima's contract)

Install

pip install "cogno-mcp[mcp]"     # [mcp] pulls the official MCP SDK (for transport)

cogno-anima (the contract) is a runtime dep not yet on PyPI — install it from git first; see .github/workflows/ci.yml. Without the [mcp] extra the MCPDispatcher still works against any session you provide; the transport helpers raise a clear MCPUnavailableError.

Use

import sys
from cogno_mcp import MCPDispatcher, stdio_session

async with stdio_session(sys.executable, args=["my_server.py"]) as session:
    dispatcher = await MCPDispatcher.create(session)          # lists + caches tools

    dispatcher.tools_schema()            # OpenAI tool defs the EGO offers
    dispatcher.is_mutating("wipe")       # from the tool's readOnlyHint
    dispatcher.requires_confirmation("wipe")   # from its destructiveHint

    await pipe.run_turn(ctx, cfg, dispatcher=dispatcher)      # cogno-soma

Transports: stdio_session(command, args=...) (subprocess), http_session(url) (Streamable HTTP), sse_session(url) (legacy SSE). Each is an async context manager that connects, initialize()s, and tears down on exit — the host owns the lifecycle.

Mapping

Skills + MCP + native together

Each source is a ToolDispatcher; merge them per persona:

from cogno_anima.tools import CompositeDispatcher
dispatcher = CompositeDispatcher([cortex_dispatcher, mcp_dispatcher, native_dispatcher])

The ToolDispatcher contract is the unifier — skill / MCP / native are just sources.

The Cogno ecosystem

cogno-mcp is one organ of Cogno — a family of small, composable, Apache-2.0 libraries that together form a complete conversational-agent platform. Each library owns a single concern and stays infra-agnostic; a host assembles them into a running agent:

The open-source libraries are the organs; the host is the body that joins them. Our reference host — cogno-host, with its cogno-ui dashboard — is the private product layer, but it holds no special powers: everything it does rides on the public seams documented in each library's docs/HOST_INTEGRATION.md, so you can assemble a body of your own.

Development

pip install -e ".[dev]"
pytest tests/unit -q            # SDK-free dispatcher logic against fakes
pytest tests/integration -q     # MCPDispatcher vs a real FastMCP server over stdio
ruff check cogno_mcp tests && mypy cogno_mcp
python examples/host_min.py     # spawns a local MCP server and calls a tool

Apache-2.0.