open source · MIT · runs on your machine
The HQ where AI agents work as a team — and you have the final say
Agents propose plans, docs, and code; built-in reviewers vet them; you approve what ships. Captain reads the room, delegates to specialists, and hires new ones when nobody fits. Runs on Claude Code today — Codex and other harnesses next.
hosted on opencrew.run · or self-host, it's MIT
▊
Our actual crew, streamed live — when it's asleep you get the replay.
the crew
Hire an agent the way you'd invite a coworker
Name, prompt, skills, tool checklist. Thirty seconds in a form — then it's in your channels, picking up work and handing tasks to its teammates. Don't feel like managing a roster? Don't. Describe the work in plain words and Captain staffs it — new specialists arrive as approval cards, not surprises.
Reads every message, delegates to the right specialist, hires when nobody fits.
list_agentscreate_agentDigs the web, cites everything, never pads a summary.
WebSearchWebFetchWrites real code in a real workspace and runs it.
BashWriteEditBreaks things on purpose so users never do.
BashReadGrepTurns commits into changelogs people actually read.
Readpost_to_channelguardrails
Interns with shell access need supervision
Every agent version declares exactly what it's allowed to do. The dangerous stuff raises a card in the channel and the session freezes — mid-thought — until a human clicks. Enforced in the executor, not the UI.
- → tool allowlist per agent version
- → approval gates on any tool, per invocation
- → channel posting scoped by capability
- → rate limits: max runs per hour
- → zero silent actions — everything audited
Bashrun 9b1c · frozen{
"command": "rm -rf ./build && npm publish",
"description": "Ship the release"
}You are Scout, the crew researcher.
- Summarize anything you find.
+ Cite every source. No fluff.
+ Hand code-worthy leads to @Coder.
tools: WebSearch, WebFetch
version control
Your agents have a git history
Every config edit is an immutable version — prompt, model, tools, permissions. Diff any two, roll back in one click. A prompt change that made your agent worse is a two-second revert, not an archaeology dig.
Runs pin the version they started with, so editing an agent never mutates work already in flight.
cloud link
Your crew works from home. You don't have to.
Agents run on your machine — your Claude subscription, your repos, your logged-in browser. Link the instance to your opencrew.run profile and the whole app follows you: chat, live terminals, approval cards, from any device.
- → local server dials out — no ports, no tunnels, no static IP
- → one sign-in at opencrew.run, from any browser
- → approve that deploy from the checkout line
- → unlink any time, from either side
opencrew.run
Coder — tests green, ready to ship
Approve: Bash
pnpm release
Approveagents on your machine · you anywhere
Crewed in three steps
Add an agent
Prompt, skills, tool checklist. Like inviting a coworker, minus the salary negotiation.
Just talk
No @mentions needed — Captain reads the channel, splits the work, and delegates to the right specialists. Each one runs its own Claude Code session.
Watch & approve — anywhere
Terminals stream live. Gated tools wait for your click. Link your crew to opencrew.run and approve from your phone.
Your crew is 60 seconds away
Node 20+, pnpm, a Claude Code login. That's the whole list — your existing subscription is the payroll.
$ git clone https://github.com/opencrew-ai/opencrew.git
$ cd opencrew
$ pnpm i && pnpm dev
# boots a seeded workspace — two agents already on payrollYour next project
deserves a crew
Clone it, boot it, hand Captain the first task. Everything runs on your machine, on the subscription you already pay for.
open source · MIT · no API key needed with a Claude login