MCP

Serve remember and query as tools of an MCP server, so MCP clients can use gligabao as their memory.

gligabao can serve remember and query as tools of an MCP (Model Context Protocol) server, so that MCP clients (Claude Desktop, Claude Code, IDE agents, …) can use it as their memory. It is built on the official Go SDK, modelcontextprotocol/go-sdk.

It is off by default and has no authentication. Enable it with mcp = true in the config file (build/gligabao.toml) or serve --mcp. Callers choose the user_id themselves, exactly as with the REST API when no Auth hook is set, so turn it on only on a trusted network. The endpoint is not covered by the Auth hook (a warning is logged at startup if both are set); when gligabao is mounted on a host’s mux, MountOptions.Wrap does cover it, as it is under /v1/gb/.

Endpoint

/v1/gb/mcp, the MCP streamable HTTP transport (POST; GET and DELETE are answered too, as the transport requires). The server is stateless: no session is kept between calls, and answers are plain JSON (no event stream).

claude mcp add --transport http gligabao http://localhost:7878/v1/gb/mcp

For a client that takes a JSON config:

{"mcpServers": {"gligabao": {"type": "http", "url": "http://localhost:7878/v1/gb/mcp"}}}

Tools

Only these two tools are exposed for now. Each runs the handler of the matching REST endpoint in process, so validation, limits, errors, monitoring events (remember.completed, query.completed, …) and results are those of the endpoint; see its page for the details of every field.

remember

Stores a text in memory, like POST /v1/gb/remember with a text.

Argument
user_id	required	The user that owns the dataset
text	required	The text to remember (at most 1 MiB)
dataset	optional	The dataset to store into, created if missing (default main)
kind	optional	knowledge (default) or episode
session_id	optional	A label for the user/agent session (Sessions)
reference_time	optional	When the text was written, RFC 3339 (default now)

The other options of the endpoint (image, models, reasoning, web_search, timeseries) are not available through MCP yet.

query

Answers a question from the graph, like POST /v1/gb/query.

Argument
user_id	required	The user that owns the dataset
query	required	The question
dataset	optional	The dataset to search (default main)
session_id	optional	A label for the user/agent session
reasoning	optional	none (default), low, medium or high

Results

On success the result is the JSON response of the endpoint, both as the tool’s structured content and as text. If the endpoint answers with an error status (invalid argument 400, dataset not found 404, no LLM key 503, failed LLM call 502, …) the result is a tool error (isError: true) whose text is the endpoint’s message and status, so that the model can see what went wrong and correct the call. Arguments that do not match the tool’s input schema (a missing user_id or text) are rejected by the MCP layer before the handler runs.

Not included yet

Authentication, the other endpoints (datasets, timeseries, ontologies) as tools, MCP resources and prompts. See the roadmap.