Engrava MCP

The Model Context Protocol server for Engrava — expose an agent memory database to any MCP client (Claude Desktop, Claude Code, Cursor, Windsurf, VS Code, …) over stdio.

engrava-mcp is a standalone, runnable package that consumes Engrava's public API. It is the one way to run Engrava as a memory server; the engrava library itself ships no MCP code.

uv tool install engrava-mcp   # recommended for daily use: a persistent install
engrava-mcp                   # spawned by your MCP client over stdio

uvx engrava-mcp runs it without an install step, and pip install engrava-mcp works too. uvx keeps its environment in a cache. When that cache is cold, the first start waits for the download; see Optional providers for the [local] extra, where the download is largest.

Installing engrava-mcp pulls in engrava transitively, so you also get the import engrava library in the same environment.

Compatibility

engrava-mcp follows Engrava's version: engrava-mcp X.Y.z targets engrava X.Y and requires engrava >=X.Y,<X.(Y+1). This is a one-way version mirror for legibility — not a lockstep: Engrava releases on its own cadence, and engrava-mcp patch releases are independent.

The dependency range is the source of truth. Normal installs resolve a compatible engrava automatically; if you pin engrava yourself, keep it within that range. If no matching engrava-mcp exists yet for an engrava newer than the table's last row, that pairing is not yet verified/supported — not broken; stay on a supported pair until a matching engrava-mcp ships.

Which package do I want?

There is no third option.

Migrating from engrava[mcp]

The server used to ship inside Engrava as the engrava[mcp] extra and an in-engrava engrava-mcp command. As of Engrava 0.5.0 it lives here instead.

• Watch out: pip install "engrava[mcp]" against Engrava 0.5 does not fail — pip ignores the now-unknown extra and quietly installs bare engrava, so it can look like the server installed when it did not. Install engrava-mcp instead.
• Update any pinned requirement strings (engrava[mcp]>=...) to depend on engrava-mcp, not just reinstall.
• Your store configuration is unchanged — the same engrava.yaml / env vars work exactly as before (see Configuration).

Configuration

The server resolves its store from environment variables, in priority order:

Recommended: give the MCP server the same engrava.yaml your application uses. The yaml is the only place to declare an embedding provider (and its model / key), which the server needs to embed a new query at search time for semantic search. With only ENGRAVA_DB_PATH set, the server emits a startup warning that semantic search is inert and points you at ENGRAVA_MCP_CONFIG.

At search time, the server embeds the query with the provider this yaml declares. Whether this server's writes embed anything is decided by the same yaml's embeddings.auto_embed, which Engrava leaves off by default. With it off, a thought created through store_thought gets no embedding, so search_memory's vector ranking cannot match it — its keyword ranking still can — and an update_thought leaves whatever embedding the thought already had as it was, not refreshed. With it on, creating a thought, or changing its essence or content, also calls the provider.

Store-hook extensions need the config path

Engrava extensions that hook the store — anything wired through an engrava.yaml's hooks: section — are attached only on the ENGRAVA_MCP_CONFIG launch. ENGRAVA_DB_PATH opens a bare database and carries no configuration, so it runs with Engrava's default hooks and cannot attach a store-hook extension. That is deliberate: it is an intentionally minimal read/write facade.

Installing such an extension and starting with ENGRAVA_DB_PATH therefore leaves its store hooks unattached in this server. When an installed package advertises any extension, the server emits a startup warning naming it — it reports what is advertised, not what each one does, since it never loads them itself — so you can tell the difference between "nothing advertised" and "advertised but nothing wired it here". If reading the installed-package metadata raises an ordinary error, the server attempts to log that instead and carries on starting. Both go through Python's logging, so whether and where they surface is up to your logging configuration. To wire a store hook, launch with ENGRAVA_MCP_CONFIG pointing at an engrava.yaml with a hooks: section:

hooks:
  class: "my_package.hooks.MyHooks"

Example engrava.yaml

database:
  path: /absolute/path/to/memory.db
embeddings:
  provider: openai-compatible # or: ollama, sentence-transformer, huggingface
  model: text-embedding-3-small
  api_key: ${OPENAI_API_KEY}

A relative database.path resolves against the server process's working directory, which the MCP client chooses, not against the yaml's folder.

Client setup

Point your MCP client at the server over stdio. For example, a typical mcp.json entry:

{
  "mcpServers": {
    "engrava": {
      "command": "engrava-mcp",
      "env": {
        "ENGRAVA_MCP_CONFIG": "/absolute/path/to/engrava.yaml"
      }
    }
  }
}

A ${VAR} value in the engrava.yaml, such as ${OPENAI_API_KEY}, is read from the server's own environment, so add that variable to the same env block, unless the client is known to pass its own environment through.

This assumes uv tool install engrava-mcp. If your client cannot find the command, give its absolute path; uv tool dir --bin prints the directory. Without an install, use "command": "uvx", "args": ["engrava-mcp"].

Use ENGRAVA_DB_PATH instead of ENGRAVA_MCP_CONFIG for the zero-config quick-start, and add "ENGRAVA_MCP_READ_ONLY": "1" for an app-writes / agent-reads deployment.

Running without uvx

engrava-mcp                  # console script
python -m engrava_mcp        # module run
python -m engrava_mcp.server # module run (server module directly)

Optional providers

For an MCP deployment, prefer an embedding provider that runs outside the server process: Ollama (provider: ollama) or an OpenAI-compatible endpoint (provider: openai-compatible). The default install already covers both, and the server then loads no embedding model itself.

The default install supports the vector backend and HTTP-based embedding providers (OpenAI / Ollama) once configured in the yaml. Heavier providers are opt-in extras that mirror Engrava's own extras:

uvx --from "engrava-mcp[local]"  engrava-mcp   # sentence-transformers (local model)
uvx --from "engrava-mcp[hf]"     engrava-mcp   # HuggingFace Inference API
uvx --from "engrava-mcp[openai]" engrava-mcp   # OpenAI-compatible embeddings deps
uvx --from "engrava-mcp[ollama]" engrava-mcp   # Ollama embeddings deps

[local] runs the model inside the server process and installs PyTorch, which can add several gigabytes. A first start on a cold uvx cache waits for that download; later starts reuse the cache. uv tool install "engrava-mcp[local]" pays the download once, at install time. A model that is not already in the local model cache is downloaded when it is first loaded.

The surface

• Tools (13): get_thought, search_memory, search_keywords, list_memory, query_memory, memory_stats, get_edges, list_edges (read); store_thought, update_thought, link_thoughts, delete_thought, delete_edge (write, gated by ENGRAVA_MCP_READ_ONLY).
• Resources (3): engrava://thought/{thought_id}, engrava://stats, engrava://recent.
• Prompts (3): summarize_recent_memory, find_related, reflect_on_topic.

query_memory accepts only MindQL FIND queries; raw SQL and every other command are rejected. It returns at most 5000 rows. A limit argument replaces the query's own LIMIT; without one, a LIMIT outside 1–5000 is refused.

get_edges traverses a thought's edges by direction (IN / OUT / BOTH); with limit, at most that many edges, the highest-weight ones first, and without it, every edge. list_edges browses edges filtered by type, source, or metadata.

link_thoughts accepts optional edge metadata (JSON fields that list_edges can filter on). search_memory accepts an optional recency_now (ISO-8601 timestamp) giving the moment to measure age against (transaction time); recency takes part in the ranking only when you pass it.

Development

pip install -e ".[dev]"
ruff check src/ tests/
ruff format --check src/ tests/
mypy --strict src/
pytest --cov --cov-fail-under=90

License

MIT