How I run support, marketing and code with a team of Claude Code and Codex agents

I run several small products on my own. For a while my setup was ten terminal tabs of Claude Code, each one halfway through something, and I could never tell which one was waiting on me. The coding worked. Everything around it (support email, follow-ups, the weekly numbers, posts) still waited until I remembered it.

So I built Brigade (brigade.lol), an open-source (MIT) app that runs a team of AI teammates. Each teammate is the real Claude Code or Codex CLI, signed in with my own Claude or ChatGPT account, with a job, its own computer, the apps I give it and a shared memory of how we work.

• Code: https://github.com/gmpetrov/brigade
• Live demo, no sign-up: https://demo.brigade.lol

Why the official CLIs and not an API

I already pay for Claude Max and ChatGPT Pro. I didn't want a second bill for tokens, and I didn't want to rebuild an agent loop that Anthropic and OpenAI improve every few weeks.

So a small runner starts the official claude or codex command on a computer, signed in with the member's account. It strips API keys from its environment, so the work runs on the subscription. When one account hits its limit, the teammate moves to the next account of the same provider and carries on in the middle of the task.

The runner is the only part that needs a computer: your own Mac or Linux box, or a cloud computer we host.

One job per teammate

My first attempt was one agent with every tool. It got vague fast, and when something went wrong I couldn't tell which instruction caused it. Now each teammate holds one job. Support answers email from the help docs. Marketing drafts posts and keeps directory listings up to date. An engineer opens pull requests, starts a live preview of the app in the thread, and asks another teammate to review before I merge.

Each one gets its own apps (Gmail, GitHub, Stripe, Notion, Linear, PostHog and others, or any MCP server), read-only or read-write, teammate by teammate. When one is off, I fix that one and leave the others alone.

I talk to them in threads, the way I would with a new hire. What we decide goes into a workspace memory and a shared library of documents, so I no longer re-explain the product, the tone and last week's decision at the start of every session.

Work that starts without me

A trigger starts a thread when something happens in an app or a webhook arrives: a failed payment, a new GitHub issue, an email. Routines run on a schedule, like the weekly report. Every step a teammate takes is logged in its thread, and there are daily caps on threads, computer hours and writes to each app.

What still comes to me

For each teammate, I choose which app actions wait for my yes: sending an email, a change in Stripe, a merge. Those requests land in one list, "Needs you", with what the teammate wants to do and why. My phone gets a push, and a tap opens the request.

This is opt-in. A teammate asks first only for the app actions I set it to ask about, and does the rest on its own. I started strict and loosened it teammate by teammate.

It builds itself

Most of the commits in the repo come from the app's own teammates, on branches named brigade/…. I make the calls that matter: what to build, what to merge, what to throw away. The merged pull requests are public if you want to see what that looks like day to day.

What doesn't work yet

• On the free plan, teammates work only while your computer is on. Close the laptop and they stop. The paid plan ($99 a month per workspace) gives them a computer that keeps going.
• It runs Claude Code and Codex only. If you want Gemini, an open model or a local one, it can't run them today.
• Teaching a task (you do it once while the teammate watches) and taking over a teammate's screen work only on the cloud computer.
• Full computer use, where a teammate reads the screen and clicks in any app, is Mac only.
• Approvals cover the app actions you pick, nothing else. A teammate with read-write GitHub access and no approval set acts on its own.
• The name is shared. Other open-source projects are called Brigade (spinabot/brigade, escoffier-labs/brigade) and have nothing to do with this one, which is Brigade (brigade.lol).

How it compares

I wrote up how Brigade (brigade.lol) compares with OpenClaw, Paperclip, Lindy, Sintra, Polsia and plain Claude Code or Codex, including where it is weaker: https://www.brigade.lol/compare

Try it

It's free on your own computer (macOS or Linux), and the whole product is in the repo.

• Live demo, no sign-up: https://demo.brigade.lol
• Source: https://github.com/gmpetrov/brigade
• Hosted version: https://www.brigade.lol

I'd like to hear where it breaks for you, here in the comments or in an issue.