Docs · MCP

Connect an agent to your brain

Cerebro speaks MCP, so any MCP-compatible agent — Claude Code, Cursor, Windsurf, or your own — can read and write your company brain. Two minutes, one endpoint, one key.

Endpoint

https://cerebro.frozo.ai/mcp

Your dashboard always shows the current endpoint at Admin → MCP.

Get an API key

Sign in as an admin or owner, open Admin → MCP, and Generate key. Name it after the agent (e.g. “Ada's Claude Code”) and copy it immediately — it's shown once.

Key format: vc_<base64url>

Wire it into Claude Code

One command adds Cerebro as an HTTP MCP server:

claude mcp add --transport http cerebro \
  https://cerebro.frozo.ai/mcp \
  --header "Authorization: Bearer vc_your_key_here"

Restart Claude Code. Ask it to “search the company brain” — it now has memory_search, memory_write, memory_context, memory_supersede, and more.

Wire it into Cursor / Windsurf / any MCP host

Add an HTTP MCP server in the host's config:

{
  "mcpServers": {
    "cerebro": {
      "url": "https://cerebro.frozo.ai/mcp",
      "headers": { "Authorization": "Bearer vc_your_key_here" }
    }
  }
}

Reload the window. The brain shows up as a tool source the agent can call.

Prove it recalls

Ask the agent to write a decision (“remember that we chose Postgres over Pinecone”), then in a fresh session ask “what did we decide about vector search?” — it recalls it, dated and sourced. That's the whole product: an agent that remembers what your team already knows.

Docs · Fleet demo

See two agents share one brain (2 minutes)

Reproduce this with nothing but the MCP config above and one API key. Two separate agent sessions — no shared context window — and the second one answers from the first one's memory.

Session A writes

Open Claude Code connected to Cerebro (steps above) and paste:

Record this in Cerebro: we decided to use edge functions for all connectors, because cold-start cost beats maintaining a worker fleet. bucket=decision, project=<your-project>.

Session B recalls

Open a fresh Claude Code session — new context window, same MCP config — and paste:

Before we touch the connector code: check Cerebro — have we made any decisions about how connectors are deployed?

What you just saw

Session B answered from session A's write — dated, and sourced with a memory ID. That's recall across agents with provenance, which per-agent memory SDKs don't do. Open Admin → AI agents and you'll see each agent's write count and last write, so you always know who taught the brain what.

[MEDIA: side-by-side sessions]

Docs · Compiler

Compile your brain into AGENTS.md

Your agents can generate their own repo-local context file from the team brain. One tool call compiles a project's active decisions, learnings, and entities (confidence ≥ 0.7, newest first) into ready-to-commit markdown — every entry dated and linked to its memory ID.

compile_context_file({ project: "your-project", format: "agents" })

Pass format: "claude" for a CLAUDE.md instead — same body, different header. Prefer clicking? Open Admin → Projects and hit Export AGENTS.md on any project row.

Every compiled file ends with a provenance footer — how many memories, compiled when, from which project — so a stale file is visibly stale.

Want the local-first core?

The MIT engine runs on your machine, storing memories as plain markdown.

Get the OSS →