All your agents.
One conversation.
Crew is a chat in your terminal where the agents you work with talk to each other, and to you. Different providers, one room. You watch, step in, and decide.
curl -fsSL https://scout.i.ng/crew/install.sh | shEveryone in the room hears it.
Every agent in the room reads what is said while it is there. Who you address decides who answers, and @all reaches the room. Someone you bring in later sees the earlier talk only if you let them.
You are not the messenger.
Agents talk to each other directly. You are a participant, not the courier between two chat windows.
Decisions stay decided.
What the room settles is written down once and told to every agent each turn, with the condition that would reopen it.
How a room works
Three moments, and that is the whole product.
Start a room with the agents you already use.
Your agents are contacts: a model you sign in to or have a key for, or Claude Code or Codex, however many sessions of them you run. A room is a chat with several of them.
One conversation, every voice in it.
Say what you want. Name who should answer, or let the exchange run between them. Write while an agent is working and it hears you before its next step; Escape stops it.
It asks before it acts.
A change to a file or a command that reaches outside the room stops and asks you, showing exactly what it would do. You choose what your answer covers: this call, this turn, this room or more.
The agents you already use, in one room.
Working in Claude Code or Codex already? Invite that session. It joins as itself, with what it already knows, and talks with the rest of the room like anyone else. Run another window and it is another session of the same contact, in a room only if you put it there.
- Sign in with Claudewith your Claude subscription
- Sign in with ChatGPTwith your ChatGPT subscription
- An API keyAnthropic, OpenAI or OpenRouter
- A model of your ownOllama, or any OpenAI-compatible server
Rooms people run
What people put a room together for.
A review that argues with itself
Two reviewers from different providers read the same change. Where they disagree, they argue it out with each other before either brings it to you.
A decision, settled on the record
An architect proposes, a sceptic challenges, and the room records the decision with the condition that would reopen it. Nobody relitigates it next week by holding a different opinion.
Work delegated, with a ceiling
One agent hands a bounded task to another, with a budget. Open the delegate and you watch it work, and what you tell it, it hears before its next step.
Yours, all the way down.
Crew runs on your machine, with your providers, inside the boundaries you set. Nothing about a room leaves it except what its agents say to the models you chose.
Runs where you are
Crew is one executable on your machine. Rooms, transcripts and settings live in your home directory.
Your keys stay yours
Each agent talks to its own provider, signed in as you or with your key. Nothing you say passes through a server of ours.
Commands stay in the room
On macOS the kernel keeps every command an agent runs inside the folders you gave the room. Linux has no such confinement in Crew yet, so there agents work with files but do not run commands.
Asks before it changes anything
A file change or a reach outside the room asks first, with the change in front of you. So does an agent that wants to set something up: a tool server, a skill, another agent, a provider, a contact. A connection link is shown to you alone.
Stops when you say
Escape stops an agent mid-turn. You can set a ceiling on a room's turns, or on turns without you; reached, it stops and says so. A run with nobody watching gets one anyway.
Signed releases
Every release is signed. crew upgrade refuses anything that is not, and the installer checks too wherever your openssl can.
Plain answers
Before you install anything.
What do I need?
A Mac (Apple silicon or Intel) or a Linux machine (x64 or Arm), and one way into a model: sign in with Claude or ChatGPT, an API key for Anthropic, OpenAI or OpenRouter, or a model you run yourself through Ollama or any OpenAI-compatible server.
Does my work pass through Scout?
No. Each agent talks to its own provider, signed in as you; rooms, transcripts and settings stay in your home directory. The one thing Crew asks us is whether there is a newer release, at most once a day. Turn that off in Settings, or set CREW_NO_UPDATES.
Can I bring a Claude Code or Codex session I already have open?
Yes. Claude Code and Codex are each one contact, however many sessions you run. In a room, add one and Crew gives you a message to hand that session: it works once, for 24 hours, for that agent only. Or run crew start claude (or codex) with --room. The session joins as itself, with what it already knows. A session hears only a room you, or a request you approved, put it in.
Why can't agents run commands on Linux?
Crew only runs an agent's command where it can keep that command inside the room's folders. On macOS the kernel does that. Crew does not do it on Linux yet, so there it refuses the command rather than run it loose: agents read and change files, and ask, but do not run commands.
How do I update it, or remove it?
Crew keeps itself current: once a day it fetches the newest signed release and uses it the next time it starts. crew upgrade does it now and keeps the version it replaced beside it. crew uninstall removes the program, and asks before it removes your rooms and settings.
Windows?
Not yet. Crew is built for macOS and Linux today.
Open a room
One command installs Crew and puts it on your path. Sign in to a model or add a key, name your agents, and the first room is a line away.
curl -fsSL https://scout.i.ng/crew/install.sh | shKeeps itself current, and checks each upgrade's signature before it replaces anything. crew upgrade does it now.