Give your Claude subscription a team.
OpenCrew turns your Claude Code subscription into a local, reviewable team of AI agents. One ask becomes a crew workflow you can watch: a Captain delegates, workers build in their own checkouts, reviewers vet, you approve. Every approval is a commit in your repo.
Quickstart · Ten minutes · Why not more terminals? · In depth · Discord · opencrew.run
Who it's for. Developers already on Claude Code who want parallel work without babysitting
three terminals. What changes. One message becomes a visible workflow with names, threads,
diffs, and a decision. Why it's safe. It runs on your laptop on the plan you already have,
no API key. Gated tools stop at an approval card. Agents never git commit. What you approve
lands in your repo, so the record is yours, not ours.
This is the run in the GIF above, typed into a project's #general:
Add a dark mode toggle to the Settings page. Spawn a frontend worker to build it and a qa worker to verify it in Chrome before it is proposed. Keep it small.
What happened, unedited: the Captain spawned frontend-1 in its own checkout of the repo. It
built the toggle, started the dev server on its reserved port, opened the page in the user's
own Chrome, clicked the toggle, reloaded to confirm it persisted, and proposed the change.
CodeReviewer read the diff and cleared it. One click on Approve & commit made commit
ca3b5df in the repo, with the decision logged in .opencrew/decisions.md. Ask to approve:
under four minutes, about $0.90 of the subscription's usage.
Prerequisites: macOS or Linux, a Claude subscription with the claude CLI logged in (the
script installs Node 22 and pnpm if missing), and about two minutes.
curl -fsSL https://opencrew.run/install | bashIt clones into ~/opencrew, starts the app on the first free ports, and opens your browser
already signed in. Then:
- What are you building? Name it. Point at its repo, or leave it empty and OpenCrew keeps a repo for it.
- Just type in the project's
#general. The Captain 🧭 reads every message and spawns workers (frontend-1,qa-1, …) for real work. Click terminal on any reply to watch. - Needs You fills up as work comes back, already reviewed by the built-in 📚 Librarian and 🔍 CodeReviewer. One click approves, and the approval is a commit.
Something off? docs/TROUBLESHOOTING.md covers the five things
that go wrong in the first ten minutes. Re-run the install line to update; remove it with
… | bash -s -- --uninstall. Codespaces and manual install are in
docs/GUIDE.md.
- Explore a codebase before a feature. "Three attempts at the checkout refactor, pick the best." Three workers, three checkouts, one CodeReviewer verdict, one diff in your inbox.
- Review a PR with specialists. Security to one worker, test coverage to another, a consolidated proposal from the Captain. Nothing merges until you say so.
- Turn a vague ask into a plan. "What would you build first here?" A researcher digs, the Captain proposes a plan doc, the Librarian checks it isn't a duplicate, you approve, and the tasks land on a board the crew works top-down.
OpenCrew keeps no agent memory of its own. What the crew proposes and you approve becomes files and commits in your repo:
.opencrew/
plans/…, notes/… docs the crew proposed and you approved, one file each
decisions.md one line per approval, rejection, or change request
The review thread rides along as a git note. After a real run, in a plain terminal:
$ git log --notes=opencrew -1
b05ebb2 docs: Shop Launch Checklist (v1)
Notes (opencrew):
OpenCrew review · "Shop Launch Checklist" v1 (plan)
Proposed by agent Captain; approved by anup-singhai.
Review comments:
- anup-singhai: Good. Keep step three about the announcement.
Agents read the same files you do. Push the repo (git push origin main refs/notes/opencrew)
and the memory goes with it. Ask a project's Captain "what did we decide about the launch?"
and the answer cites path@sha. Leave OpenCrew any time; the record stays.
Most of what a crew decides is small: is this message an ask or a thank-you, is git status
safe to run, is this plan complete enough to show you. OpenCrew answers those with a typed
judgment (about 100 ms, TypeSafe's Jev) instead of a whole Claude Code turn: an ack costs
nothing, a lane-obvious ask goes straight to the specialist, simple chat gets a fast model,
read-only commands stop waiting on an approval card, and obvious docs skip the reviewer.
When a judgment is unsure or unavailable, the crew does exactly what it did before. Every
judgment is logged in the run's terminal panel; each one can be switched off in Settings →
Fast judgments, which also says what text each decision sends. The reply streams token by
token, and the Speed panel on Today shows p50/p95 per phase so you can see the difference.
Two more levers, both optional. Settings → Engine switches agent turns from Claude Code
(your own login, a process per turn) to OpenCrew's engine: the Claude API called directly on
your key or on OpenCrew credit, with a hot prompt cache, tools running in parallel, and no
process to boot — chat-class turns can run on Wafer's fast open models when you add a Wafer
key. And install.sh sets up a local code + record index (CocoIndex Code) so agents retrieve
the whole function or doc section that matters instead of reading files at model speed; the
search tool exposes the same index to them. Both fall back to today's behaviour when absent.
| Three Claude Code tabs | OpenCrew | |
|---|---|---|
| Who coordinates | You, from memory | A Captain per project; workers report in threads |
| What you see | Three scrollbacks | Rooms, live terminals, diffs, one inbox |
| Same repo, same time | Edits collide | Each worker has its own worktree and port |
| Risky commands | Whatever the tab does | Stop at a card; approve from your phone |
| What survives | Nothing | Every approval is a commit; the review is a git note |
| Where it runs | Your laptop | Your laptop, same subscription, no API key |
When not to use it: a one-shot task you'd finish in one prompt; a repo you can't let anyone commit to; a team with no approval habit yet. What's still early: no mid-turn steering, no DMs or push notifications, Linux and macOS only. The rest of the rough edges are in docs/GUIDE.md.
Agents and repos stay on your machine. In ⚙ Settings, link the crew to
opencrew.run and open the same app from your phone or any device;
sign in once. The crew is your laptop, so it has to stay awake and running for the phone
to work: keep it plugged in with sleep off (on a Mac, caffeinate -i in a terminal, or
Battery → Options → prevent sleeping on power). Prefer your own front door? A Cloudflare
tunnel works too (OPENCREW_TUNNEL_TOKEN and OPENCREW_TUNNEL_URL, see
docs/GUIDE.md).
Everything on your laptop is free, forever. No seat limit, no agent limit, nothing local is ever gated, and the record in your repo is yours whether you pay or leave.
Pro is the one thing the laptop cannot do alone: decide from anywhere. When your crew finishes something while you are away, you open the same app on your phone through opencrew.run and press approve; your laptop makes the commit. That press is a remote decision (approve, reject, send back, answer a tool approval).
| Free | Pro · $19/month | |
|---|---|---|
| Everything on the laptop | ✓ | ✓ |
| Remote decisions via opencrew.run | 3 a month | Unlimited |
| Teammates with their own login | ✓ | |
| Share a change with someone who has no install | ✓ | |
| Judge credit (fast judgments through opencrew.run) | $5 once | +$5 a month |
| Inference credit (the OpenCrew engine without your own API key) | $5 once | +$10 a month |
Self-host the relay and there is no plan at all. Why the line is here: Free on your laptop, Pro in your pocket.
Environment variables or a .env at the repo root. Everything has a working default.
| Variable | Default | What it does |
|---|---|---|
PORT / OPENCREW_WEB_PORT |
3001 / 5173 |
API and web ports |
DATABASE_URL |
data/opencrew.pgdata |
Embedded PGlite path, or a Postgres URL |
OPENCREW_REPOS |
data/repos |
Repos OpenCrew keeps itself (folderless projects, HQ) |
OPENCREW_PROJECTS_INDEX |
data/projects.json |
Which folders were projects; a reset database offers them back |
OPENCREW_ENVS |
data/envs |
Per-agent worktrees of project repos |
OPENCREW_CONCURRENCY |
8 |
Concurrent agent turns (2 reserved for you) |
OPENCREW_LOCAL_AUTOLOGIN |
1 |
Browser on this machine is signed in; 0 requires the form |
OPENCREW_TELEMETRY |
1 |
Anonymous daily ping: counts and versions only, never content (source). 0 turns it off; also a toggle in Settings |
OPENCREW_RELAY_URL |
https://relay.opencrew.run |
Cloud Link relay, self-hostable |
OPENCREW_CLOUD_MODE |
unset | 1 = a hosted crew (the Dockerfile): binds every interface, no auto sign-in, no telemetry, needs SESSION_SECRET, links itself from OPENCREW_CLOUD_WORKSPACE_ID + OPENCREW_CLOUD_LINK_SECRET (minted by the relay) |
OPENCREW_WAFER_WORK_MODEL / _CHAT_MODEL |
GLM-5.3 / DeepSeek-V4.1-Flash |
The fast tier's models for work and chat-class turns (source) |
OPENCREW_JUDGE |
on |
off disables every fast judgment (a linked crew judges through opencrew.run) |
OPENCREW_INDEX |
1 |
0 turns off the code + record index (CocoIndex Code, installed by install.sh); agents then read files instead of retrieving whole functions and doc sections |
TYPESAFE_API_KEY |
(unset) | Self-hosted: judge directly with your own TypeSafe key |
ANTHROPIC_API_KEY |
(unset) | Key for the OpenCrew engine (Settings → Engine stores one encrypted, too); Claude Code turns use your claude login |
WAFER_API_KEY |
(unset) | Optional: Wafer's fast open models for chat-class turns |
More in apps/server/src/env.ts.
pnpm dev # web :5173 + server :3001, server hot-reload
pnpm test # Vitest: fabric, guardrails, record, ask, environments…
pnpm build # type-check and build everythingCONTRIBUTING.md has the ground rules (the guardrail choke points are sacred); docs/GUIDE.md shows how to add a tool in one file. Did your crew do something worth showing? Open a Share your workflow issue; the best ones get featured here. MIT licensed.
If OpenCrew saved you time, star the repo. It helps other Claude Code users find it.