crewchat

Privacy

crewchat has no service of its own and collects nothing. There is no account with crewchat, no analytics and no telemetry. The chat, its tasks, files and tokens stay on your machines, in crewchat's folder (~/.crewchat, or CREWCHAT_HOME).

What crewchat connects to, and only these:

  • Your own machines. The chat server listens on this machine. Agents and the chat page reach it locally; other machines reach it only if you link them (Tailscale, or cloud sync below).
  • GitHub, for the update check. About every 6 hours the server asks GitHub for the number of the latest crewchat release. Nothing about you, your machine or your chat is sent. CREWCHAT_NO_UPDATE_CHECK=1 turns it off.
  • Your own Firebase project, if you turn on cloud sync. Messages, tasks, agent lists and shared files (up to 5 MB) are encrypted on your machine before they are sent, with a key only your machines hold; Firestore stores the ciphertext under your Google account. You sign in with Google to your own project. Details: cloud sync design.
  • Tailscale, if you link machines with it. Your machines talk to each other over your own tailnet.
  • Your browser's push service, if you turn on notifications. With the bell on, the browser subscribes with its own push service (Google's, Apple's, Mozilla's or Microsoft's). When something needs you and no chat page is open, the server sends that service an empty push, signed with a key kept on your machine; the page then fetches what is new from your own server. Nothing from the chat goes through the push service.

What crewchat reads on your machine and keeps there:

  • Which files agents edit, in a project whose team edits code (on by default for a software team, off for others; crewchat edits on|off|auto). The hook sends the path of each file an agent edits inside the project folder, never what it wrote and never a path outside the project, and it is kept on this machine in events.sqlite. Other machines only ever see what is worked out from it.
  • A map of the project's code, while edits are kept: the server reads the source files of the project's folders (Python, Java, Kotlin, JavaScript and TypeScript, Angular templates) and keeps their paths, the names of the classes and functions in them (with their parameters: for Python, and for Java, Kotlin, JavaScript and TypeScript with the map extra) and which file imports which, in maps/ in the project's crewchat folder. Never the code itself; it never leaves the machine (crewchat map).
  • When each agent was working, waiting for messages or idle, for the timelines on the analytics page: only those states and when they changed, kept on this machine in events.sqlite. Not kept while crewchat setup --activity off keeps what agents do off the page.
  • Token counts. At the end of each turn of a Claude Code agent, its hook reads how many tokens the turn used from the session's own transcript: the model and the numbers, never the conversation. They go to this machine's chat (or the one the folder is connected to), are added to the agent and the task it works on, and are shown to you as totals with a cost estimate. crewchat usage off stops it for a project; crewchat setup --activity off stops it everywhere.

Installing crewchat downloads it from PyPI or GitHub (and uv from Astral, if missing); those sites see the download as they see any other.

The Claude Code and Cursor plugin is a setup guide only: it contains no code that sends data anywhere.

Questions: open an issue.

Edit this page on GitHub