Skip to directory

Cognee MCP

Cognee (official)

stdio by default; documented SSE and Streamable HTTP modes5+ documented tools1

About Cognee MCP

Cognee's official MCP server exposes a focused memory API for storing, recalling, and forgetting agent memory, with optional code-graph search and other tools discoverable through its tool-search mode.

What you can use it for

  • Persistent agent memory
  • Session memory
  • Code graph search
Connection and setup
Hosting
Self-hosted local process; optional Cognee Cloud or existing Cognee API mode
Transport
stdio by default; documented SSE and Streamable HTTP modes
Endpoint
Local process; no hosted endpoint documented
Authentication
Local direct mode has no separate MCP authentication, but the default model configuration needs an LLM API key. Cloud mode uses COGNEE_API_KEY; API mode may require API_TOKEN when the API is protected.

Requirements

  • Python environment and uv package manager; the README does not state a minimum Python version.
  • For default local model use, an OpenAI API key or another supported model/embedding configuration; host sampling can replace the LLM key only when supported, while embeddings still need a provider.

Setup steps

  1. Clone https://github.com/topoteretes/cognee and enter its cognee-mcp directory.
  2. Install uv if needed, then run `uv sync --dev --all-extras --reinstall`.
  3. For the documented default setup, create the MCP directory .env file with `LLM_API_KEY=YOUR_OPENAI_API_KEY`; obtain an OpenAI key at https://platform.openai.com/api-keys. Use the official environment template to configure another supported provider.
  4. Run `python src/server.py` for stdio and configure the MCP client to launch that command with the cognee-mcp directory as its working directory.
  5. Reconnect the client and inspect tools/list. The default pinned list includes remember, recall, and forget; other tools are discoverable via search_tools.
Tools 5+

The official README says three tools are pinned by default; other tools are discovered dynamically. `cognify_status` is documented but unadvertised by default. Inventory is not exact for the default tools/list response.

  • remember

    Store text or file content as session or persistent memory.

    01
  • recall

    Search stored memory.

    02
  • forget

    Delete a selected dataset, item, or owned memory.

    03
  • cognify_status

    Check background ingestion progress; documented as discoverable rather than pinned.

    04
  • code_search

    Search indexed code graphs using documented graph operations.

    05

This server’s available tools can vary by account, permissions, or configuration. The list shown is not its full runtime catalog.

Limitations and notes
  • Default direct setup requires a model key and suitable embedding provider unless host sampling is supported for completions.
  • Tools are dynamically exposed; the README documents a default pinned subset and discoverable tools, not a fixed complete inventory.
  • No MCP initialize, tools/list, or representative tool call was performed for this catalog entry.
Original sources
  • official vendor documentation

    Cognee docs describe its MCP memory and code-search capabilities; available tools depend on server version/configuration.

Documentation checked 2026-10-28T00:00:00Z. Connection has not been runtime tested.

Similar tools