Documentation

Cammut documentation

Cammut is a personal MCP gateway. You register upstream MCP servers once, give each agent a scoped token, and every agent reaches all of it through a single endpoint. On top of the gateway it adds agent memory (notes, a project brief, recall, handoff) and an optional encrypted archive of full transcripts.

text

 Claude Code ─┐                       ┌─► upstream MCP servers (HTTP)
 Codex CLI  ──┼─ Bearer token ─► /mcp │
 Warp       ──┘     Cammut ───────────┼─► memory      (in-process: notes, brief, recall, handoff)
                       │              └─► console     (/console/, password + authenticator)
                       └─► archive: age ciphertext only; the server cannot decrypt it

Where to start #

You want to	Read
Run it and connect one agent	Quickstart, then the Setup guide
Understand tokens, scopes and profiles	Concepts
Configure a server for real use	Install and configure
Configure Claude Code, Codex CLI or Warp	Agents
Use memory across agents	Memory
Keep full transcripts safely	Encrypted archive
Know what is protected and what is not	Security and Limits

Design in four sentences #

One door. Agents see one URL and four tools, however many servers sit behind it.
Least privilege. A token carries scopes, an optional profile, and can be revoked on its own.
Small. One static Go binary, one SQLite file, about 17 MB resident when idle.
Fails open. If Cammut is unreachable your agent keeps working; hooks log one line and move on.

Who it is for #

One operator running a few agents on a small server. The design accepts that scopes have no project dimension, which is fine for one person and a reason to think twice before handing a write token to a less-trusted agent. See Limits.