MCP overview
Use Breeth from any MCP-compatible AI client.
Breeth ships an MCP server at https://mcp.thebreeth.com/mcp. Any client that speaks the Model Context Protocol — Claude Desktop, Cursor, others — can connect to it, present Breeth's tools to the underlying agent, and let the agent read and write memory directly during a session.
What MCP does for you
Without MCP, integrating memory means writing custom adapters per client — call the REST API yourself, format the responses, plumb context. With MCP, you drop a configuration block into the client and the agent gets four tools out of the box:
| Tool | What it does |
|---|---|
add_episode | Write a prose memory |
record_fact | Write a single S-P-O triple |
search | Retrieve facts by query |
retract | Soft-delete an edge (admin keys only) |
The agent calls these like any other tool. You don't have to teach it the REST shape.
Authentication
The MCP server authenticates with the same ck_live_… keys as the REST API. You paste the key into the client's MCP server config; the server forwards it to cogram-core on every tool call.
Scopes and project bindings work the same way as REST. A write-only key can call add_episode and record_fact but not retract.
Multi-team keys
If your key is bound to multiple teams, the MCP server forwards an X-Cogram-Team-Id header on each request to pick the active team. Clients that support custom headers can set this themselves; otherwise the key picks its first team by default.
Compatibility
Tested with:
The transport is plain HTTP+JSON per the MCP spec, so any compliant client should work. If yours doesn't, tell us — we'll work through it.
Versus REST
Use MCP when an AI agent should reach into Breeth during its own reasoning loop. Use REST for everything else: server-to-server ingestion, batch backfills, your own dashboards.
The two surfaces share auth, scopes, and the underlying graph. A memory written via MCP is identical to one written via REST.