A well-scoped agent ships the feature you asked for. No drift, no AI slop.
Treelab installs, in one command, a complete method for Claude Code and Codex inside cmux: tight scoping, one objective per lot, evidence before delivery, and as many parallel worktrees as tasks.
curl -fsSL https://treelab.sh/install | bash
Then lab setup checks your Mac, asks for your key and prepares cmux.
Illustration: three projects, their worktrees and each one’s agent in a cmux window. Fictional names; real commands, simplified output (the CLI speaks French).
Four commands, from task to merged code.
Each step is a lab command. The agent does the rest in its cmux tab: it interviews you, sets its objective, tests, then hands you the next command.
All four steps stay readable without animation.
Step 1
Install
One command installs the lab CLI, the skills, the hooks and the review agents. lab setup checks your Mac, asks for your key, backs up what it changes and waits for your approval.
Illustration
$curl -fsSL https://treelab.sh/install | bashTreelab installé dans ~/.lab$lab setup✓macOS · git · jq · cmux · Claude Code · Codex✓Clé de licence acceptée✓Sauvegardes : CLAUDE.md, AGENTS.md, settings.jsonSaisis j’accepte pour continuer : j’accepte✓lab doctor : 0 rouge
Step 2
Open and scope a worktree
One task, one worktree: lab new creates the branch, the isolated folder and the cmux tab, then starts Claude Code or Codex. The agent starts by interviewing you (grill-me, ten questions at most), then splits the task into lots with completion criteria. Nothing vague gets started.
Illustration
$lab new boutique checkout --agent claudeWORKTREE=.worktrees/checkoutBRANCH=feat/checkout✓onglet cmux ouvert · Claude Code lancé▸grill-me : 8 questions, 2 relances✓treeplan : 3 lots, critères de fin écritslot-01-panier-express.md prêt
Step 3
Let the agent work
Every lot starts from a /goal built automatically from its document: what to deliver and how to prove it. The agent works in stages, tests and keeps the log, with no detours and no random code. Come back whenever you want: lab next resumes the session with its context intact.
Illustration
$lab next boutique checkout/goal reconstruit depuis lot-02-paiement-express.mdcritères de fin : 4 · preuves : tests, revue▸étape 1/3 · formulaire de carte✓12 tests▸étape 2/3 · validation serveur✓22 testsjournal et Reprise mis à jour · commit (lot-02)
Step 4
Deliver, without memorising a command
An independent review, passing checks and your approval come before the merge. At the end of every lot, the agent writes its handoff and gives you the single next command to type. lab close removes a worktree only with your explicit consent.
Illustration
$lab merge boutique checkout✓revue indépendante : 0 écart✓contrôles GitHub : verts✓PR #42 fusionnée dans mainReprise enregistrée · lot suivant : notifications▸À taper maintenant : lab next boutique checkout
As many worktrees as tasks. A scoped agent in each one.
Each worktree has its branch, its cmux tab and its agent. Switch between them; lab status shows where each one stands.
Diagram: three projects, five worktrees, each with its agent, rejoining their main branch once delivered.
Scoped from the first minute
When a worktree opens, grill-me asks ten questions at most, then the task is split into lots with completion criteria. No vague feature, no drift.
One objective per lot, built for you
The /goal is rebuilt from the lot document: what to deliver and how to prove it. The agent goes straight to the point and ships the requested feature.
No command to memorise
At the end of every lot, the agent writes its handoff and prints the single next command. You type it; the session resumes with its context intact.
Coming
Artifacts and a front-end skill, on Codex as on Claude Code
Built-in visual previews and a front-end design skill, so Codex designs interfaces as well as Claude Code does. Under construction, no announced date.
One payment. One licence. Lifetime updates.
Both offers include the same rights. The founder pack is limited to 20 copies. Prices exclude tax: the final amount, currency and taxes are shown before payment.
No. Treelab organises Claude Code or Codex’s work: the cycle, the lab commands, the scoping and execution skills, the hooks, the review agents and the templates. Access to the agent remains separate. The CLI and its messages are in French for now.
How many projects and worktrees?
As many as you open: lab init registers each repository, lab new creates one worktree per task, lab status and lab plan show them all. One active lot and one live session per worktree.
What remains my responsibility?
Scoping, permissions and approving the result. The tests to run depend on your repository. The method and its guardrails do not guarantee defect-free code.
Can I go back?
lab update keeps the previous version and --rollback returns to it. lab uninstall replays its backups. lab doctor --report prepares an anonymised diagnostic without sending anything; support is by email, with no guaranteed response time.
What happens after the 20 packs?
The founder offer closes. The standard licence remains a separate offer: an accepted price is never silently replaced. Availability is shared across languages and countries.