What it is: A liaison agent. Instead of juggling many agent sessions yourself, you talk to one agent (the first mate) that dispatches and supervises a crew of worker agents. It's the capstone of the workflow — all the other tools wired into one. Free & OSS, and new.
- Decomposes a request into parallel tasks (e.g. "add an update command to all 3 CLIs" → 3 tasks).
- Spins up each in its own tmux window with a clean treehouse worktree.
- Runs a worker agent per task, then runs no-mistakes to validate → PR.
- Two task types: ship (produces PRs / local merges) and scout (investigation reports). The first mate stays read-only on your projects except safe ops + approved merges.
- You just keep giving it work and asking questions; it does the context-switching.
./setup.sh firstmate- Clones the repo to
~/.local/share/firstmate(kept up to date on re-run). - Adds a
firstmateshell launcher (via~/.config/agentic-dev-setup.zsh) that runs your agent harness inside that clone — which is how firstmate is meant to run.
firstmate # launches Claude Code inside the firstmate repo
FIRSTMATE_AGENT=codex firstmate # or use a different harnessThen just talk to it:
- "Work on lavish-axi, gh-axi, and chrome-devtools-axi — add an
updatecommand to each CLI that bumps to the latest npm version." → it runs 3 parallel ship tasks. - "Look at the 3 most recent open issues in lavish-axi and tell me which are actionable." → a scout task, while the ship tasks run in the background.
On first run it does setup and asks preferences, including the per-project delivery mode: full gates to PR (uses no-mistakes), direct-PR, or local-only.
- Run it inside tmux — each crewmate appears in its own tmux window; that's the intended experience.
- It depends on the rest of the stack (tmux, git, gh, treehouse, no-mistakes). On first launch it detects missing deps and asks consent before installing.
- Because firstmate removes so much overhead, the bottleneck shifts to you — knowing what to build. That's the "captain's mindset" the video ends on.