Squadroom

Squadroom lets you put several coding agents on a job as a team. They message each other on their own, each in a role you pick from the built-in ones or write yourself as a prompt, and they only come to you when they need a decision. It's a web page you run on your own Mac, and you can have several repos open with a different team on each piece of work.

Website: squadroom.github.io

It uses your own accounts and keys, and you open it in a browser (a phone works too) at a private https address that only your GitHub account can sign in to. Inside the app, and in its settings and file names, Squadroom calls itself the Console.

Status: early, and macOS only. It works, but it isn't packaged yet. Setup takes an afternoon and needs sudo once. Linux and Windows aren't supported. There are no numbered releases yet: use the latest commit on main.

How a team works together

A team is the group of agents on one piece of work. Each of them has team tools, and with them an agent can send a teammate a message (team_send), ask a question and get the answer back (team_ask, team_reply), merge in a teammate's work or, in a GitHub project, the latest base branch from GitHub (team_pull), claim the files it's about to edit (team_claim), ask a teammate to hold off (team_hold), ask you to push its branch and open a pull request (team_push), read the review on it (team_pr), and read the checks GitHub Actions ran on it, with the end of a failed job's log (team_checks). A message waits until the other agent finishes its current turn, then arrives as its next prompt. Routes pass one agent's finished reply straight to another, for example when the Lead ends with "ready for review".

When a team has an Orchestrator you give it the goal, and it reports back to you in its window. With a pair you talk to the Lead. You can type into any agent's window at any time. Everything else that needs you goes into one tray, Needs you, which covers every project: network requests, questions, pushes, notes agents leave for you (team_report) and paused tabs. The Team panel shows every message between a tab's agents, and in an agent's window a teammate's message has its own style, so you can tell it from yours.

A message you send while an agent is working waits in its queue. For a Claude Code or Codex agent, your first queued message gets a Steer button: click it and the message goes into the turn the agent is working on, and the agent reads it at its next step. It shows in the transcript where the agent got it. Other agents can't take a message mid-turn, so theirs stay queued.

In an agent's window, two or more tool rows in a row show as one line, such as "18 steps: 10 Edit, 5 Read, 2 Write, 1 Grep". Click it to see the rows. A run with a failed tool says so and starts open.

Each website an agent wants to reach waits for you in Needs you. A Claude Code or Codex agent's window shows Network ask or Network auto in its header, and a click switches it. On Auto the agent reaches any website without asking, and each request is still noted in its window. A tab's menu has a Network auto switch for all its agents, and a network request in Needs you has Allow, Network to Auto. Turning Auto on needs a fresh e-mailed code. Read the threat model before you use it: an agent on Auto could send anything it can read to any site.

A permission request closes when its turn ends, for example at a time limit while you were away. It then stays in Needs you, marked "got no answer". You can allow it for next time (the agent is told to try again), have the agent ask again, or dismiss it.

In a GitHub project an agent's folder is filled once, when it is made. When the base branch gets new commits on GitHub afterwards, the agent's window says so ("main has new commits") and the agent is told with its next message. It brings them in with team_pull(from: "main"); nothing is merged by itself.

An agent can also propose a new tab with a team of its own (team_propose_tab), for work that is better done apart. The proposal shows up in Needs you with Create, Edit… and Decline. Nothing is made until you choose, and the agent is told what you chose. If you create the tab, its Orchestrator (or a pair's Lead) and the agent that proposed it can message each other across the two tabs, and the new tab's result is sent back when its team has finished.

Every agent works in its own Git clone on its own console/… branch (a Code analysis agent, which changes no files, sits on the base branch). Squadroom commits what changed at the end of each turn, so you can check the work one turn at a time in the window's Changes view. Files, in the same menu, lists everything in the agent's folder, Markdown documents first, and opens a file beside the conversation. Each commit's message holds your whole request for that turn, and pushing the branch publishes it. Nothing is pushed until you approve it, and when an agent opens a pull request you see its title and description before they are posted, unless you chose Always for that agent's pushes.

Roles

A role is a prompt with a name. It tells the agent what its job is and how to work with the others. There are eight built in: Orchestrator, Planner, Architect, Lead, Pair Lead, Designer, Reviewer and Tester. You pick one when you add an agent, and you can change the prompt for that one agent.

This is the beginning of the README. Read the rest at the original. Read the full README ↗