Run coding agents like an engineering team.
AgentCrew is the command center above Claude Code and Codex. Describe the outcome, approve the plan, and get back a draft pull request where every requirement carries its own evidence.
- 1PRD7 requirements, 2 open questions
- 2Plan5 tasks · dependency graph valid
- 3Run3 worktrees in parallel
- Claude Codeinvite-model
- Codexinvite-api
- Claude Codeinvite-ui
- read src/team/*
- npm test
- grep invite_
- cargo test
- vitest run
- git commit
- 4Verify7/7 requirements have fresh evidence
- 5DeliverDraft PR #412 · evidence table attached
Plan
A specification before a prompt.
Import the work, keep the source
Paste a brief, drop in a Markdown spec or pull Linear issues. AgentCrew keeps the source IDs, URLs and versions so every requirement is traceable back to who asked for it.
Requirements with stable IDs
Each requirement gets an ID and acceptance criteria. Saves create immutable revisions, so a change mid-mission is a new revision, never a silent overwrite.
A task graph that must validate
Plans cover every requirement, declare dependencies and cannot contain cycles. Invalid or stale plans are rejected before anything runs.
REQ-invite-01Owners can generate an invite linkREQ-invite-02Links expire after 7 daysREQ-invite-03Revoking a link invalidates it immediatelyREQ-invite-04Accepting joins the correct workspaceOPENShould invites be limited per plan tier?
Run
Parallel agents, without the collisions.
A Git worktree for every task
Every task runs in its own worktree with its own provider session. Agents cannot trample each other, and you can inspect any task mid-flight.
An integration branch that grows as tasks land
Finished tasks are committed and merged into the run branch so dependents start from real work. Conflicts are never auto-resolved; the task blocks until you decide.
Any runtime, per task
Assign Claude Code to the model layer and Codex to the API in the same plan. Roles stay stable while the runtime behind them changes.
Prove
“Done” means verified, not claimed.
Evidence per requirement
A verifier role exercises the behaviour and records test output, logs or reproductions against the requirement ID and the exact commit.
Stale evidence closes the gate
Change the PRD or push new commits and the evidence goes stale. Completion is refused until it is fresh again. No exceptions, no override flag.
A draft PR with the receipts
Delivery prepares a draft pull request whose body carries a per-requirement evidence table. Reviewers see what was proven, not what was promised.
| Requirement | Evidence | Commit |
|---|---|---|
| REQ-invite-01 | 12 tests pass | a41c9e |
| REQ-invite-02 | expiry fixture | a41c9e |
| REQ-invite-03 | revoke repro | b02d17 |
| REQ-invite-04 | e2e join flow | b02d17 |
Recover
Crash, sleep, quota, resume.
Automatic recovery on restart
If the engine dies mid-task, the next start checkpoints each interrupted worktree, re-queues the task in place and resumes with a continuation note. No operator ritual.
Steer a running task
Redirect a task in flight with one instruction. The provider is interrupted cleanly, the worktree checkpointed, and the task continues with your note at the top of its prompt.
Quota-aware scheduling
AgentCrew tracks session capacity and provider quota so a burst of parallel tasks never turns into a wall of rate-limit errors at 2 a.m.
- 02:14:07Task invite-ui running · Claude Code
- 02:16:31Engine process lost (laptop slept)
- 07:48:02Engine start · 1 run recovery_required
- 07:48:03Provider pid 48122 not alive · checkpointed worktree
- 07:48:04Task re-queued in same worktree · continuation note added
- 07:48:09Run resumed · 2 tasks remaining
Runtimes
Bring the agents you already pay for.
AgentCrew drives unmodified CLIs through your existing subscriptions. No token resale, no proxy keys, no new bill for model usage.
By the numbers
Four numbers that decide whether you can trust an agent with your repository.
- 1
- command carries a written intent all the way to a draft pull request. You are interrupted twice, on purpose.
- 5
- agent runtimes behind one set of crew roles, running on the Claude and OpenAI subscriptions you already pay for.
- 2
- approvals stand between a plan and your branch: one on the plan, one on delivery. Never zero, and never a merge you did not see coming.
- 0
- bytes of your code on our servers. The engine, the state and every worktree stay on your Mac.
Each figure describes how AgentCrew is built, not what we hope it will do. It is pre-release and macOS first, and support for runtimes beyond Claude Code and Codex is still marked experimental.
Who it is for
Built for people who ship without a big team.
You are the whole engineering org.
Stop being the human message bus between three agent windows. Set the outcome, approve the plan, review the evidence.
Why founders Small product teamsTwo to seven engineers, many agents.
Give everyone the same crew, the same evidence standard and the same draft-PR handoff, whatever runtime they prefer.
Why teams Multi-repo engineersWork spread across repos and trackers.
Import from Linear, keep source links, and let the Bridge carry context across projects so you stop re-explaining the codebase.
Why multi-repoPricing
Half your first year, for the first 100.
One plan, priced at launch. Every feature, every runtime, unlimited projects. Your model subscriptions are your own.
Half your first year
Everything below, in the single plan. The first 100 people on the waitlist pay half for their first twelve months.
Everything, in both plans
- Native macOS app, terminal dashboard and CLI
- Unlimited projects and missions
- Claude Code, Codex, Cursor, Gemini CLI and Droid runtimes
- Parallel tasks in isolated Git worktrees
- Evidence-gated completion and draft PR delivery
- Crash recovery and steerable tasks
- MCP servers, skills and hooks per role
- All updates during your subscription
The first 100 people on the waitlist pay half for their first twelve months. After that, launch price. Model usage is billed by your provider, never marked up by us.
FAQ
Questions people ask before they trust an agent with their repo.
Does AgentCrew replace Claude Code or Codex?
No. AgentCrew is the layer above them. It plans the work, dispatches each task to the runtime you choose (Claude Code, Codex, Cursor, Gemini CLI or Droid), isolates every task in its own Git worktree, and gates completion on verified evidence. Your existing subscriptions and logins are used as-is.
Does my code leave my machine?
Not through AgentCrew. The engine, the SQLite state and every worktree live on your Mac. The only network traffic is what your chosen agent runtime already makes to its own provider. AgentCrew has no cloud backend that sees your repository.
What does "evidence-gated" actually mean?
A mission cannot be marked complete until every requirement in its PRD has fresh verifier evidence tied to the exact commit and PRD revision. If the PRD changes or new commits land, the evidence goes stale and the gate closes again. The agent saying "done" is never enough.
What do I get for joining the AgentCrew waitlist?
First access when the macOS build ships, in the order people signed up. The first 100 people on the list also pay half for their first twelve months. You can tell us which runtimes and stack you use and we will prioritise them. We send one email when it is ready, and no drip sequence.
Which platforms are supported?
macOS first, with a native app, a terminal dashboard and a scriptable CLI that all drive the same local engine. Linux and Windows support for the engine and CLI are on the roadmap after the macOS release.
Will it auto-merge to main?
No, and it is not a setting. Delivery prepares a draft pull request with a per-requirement evidence table. Merging is your decision, made in your Git host, every time.
Founding crew open
One hundred founding seats.
Join the waitlist and your first twelve months cost half. One email when AgentCrew is ready for your Mac, and nothing else.