How Chain talks to Claude Code
No fork, no wrapper around the model. Chain uses the two things Claude Code already supports: hooks and tools. When you start a session through Chain, it has the board tools (board, claim, release, log, done) and a pre-tool hook that runs before every edit. The hook refuses writes to paths another agent has claimed or that you have blocked, and tells Claude who has it and why.
One button, or one command
Press Open AI in the app and pick Claude Code, or run it from any terminal:
$ chain open claude -t "Password reset flow"
Either way Claude starts in a fresh worktree on a branch named {user}/{task-slug}, with the rules and hook written into the repo between chain:begin / chain:end markers.
Worktrees are the start, not the finish
Claude Code can already run in its own worktree. That stops two sessions from overwriting the same files on disk. It does not stop them from both rewriting the same function on different branches, and it does nothing for a teammate’s session on another Mac. Chain adds the shared board and the hook on top, so the worktree isolation turns into real coordination. The guide to running multiple agents on one repo covers the whole chain.
Your keys stay in your Keychain
Every Claude Code session runs on its owner’s Mac, with its owner’s Anthropic subscription or API key. Teammates join a co-work with a code, never with your credentials. The relay carries claims, tasks and log lines. Your code and your terminal output stay on your screen.
Mixed teams are fine
One teammate on Claude Code, one on Codex CLI, one on Cursor: all three read the same board and respect the same claims. Chain also works with OpenClaw and any agent that speaks MCP.