Developer

How Outworked Runs Claude Code Agents as a Team in a Pixel Office

· Source: the web

Quick answer

Outworked is a free, open source macOS app where Claude Code agents work as a team inside a pixel-art office. You describe a goal, and an orchestrator splits it into subtasks and routes them to agents you have set up with a name, role, model and sprite.

What they built

Outworked is a desktop app for macOS, published on GitHub by georgesayoub, that puts Claude Code agents to work as a team inside a pixel-art office. According to the 0.3.0 release page, you give each agent a name, a role, a model and a sprite, then describe a goal. An orchestrator breaks the goal into subtasks and routes them to the right agents. The agents walk to their desks, and their coding, browsing and messaging is visible in the office as it happens.

The page says the app is free and open source under GPL-3.0. You need Claude Code installed and authenticated to use it.

How it works

The release page describes these pieces:

  • Agents: each one has full Claude Code tool access and a persistent session. Each agent can use a different model, for example Opus for hard problems and Sonnet for fast ones.
  • Teamwork: several agents can work on subtasks at the same time, and they can ask each other questions over a shared message bus.
  • Browser: a built-in Chromium browser. When a site needs a login, a window opens so you can sign in yourself, and the agent carries on with that session.
  • Tunneling: built-in Cloudflare tunnels so an agent can expose a local server with one tool call.
  • Scheduling: tasks can run on a cron, on an interval, once, or in response to an incoming message.
  • Messaging: iMessage works on macOS for sending and receiving. Slack is listed for the next release.
  • Skills and MCP: skills are SKILL.md files, with Web Browser and Scheduler bundled. There are MCP presets for GitHub, Filesystem, PostgreSQL, Slack, Linear, Playwright, Memory and Fetch.
  • Safety and cost: an approval prompt for sensitive tool calls, command allowlists for each agent, a permissions dashboard, and a cost dashboard that tracks tokens and spend per agent, session and day. The page notes you can turn the approvals off with one button.

The page gives example workflows, not reports from real users. In one, a Frontend Dev agent builds a landing page, a Designer agent reviews it in the browser, and the link goes out by iMessage through a tunnel. In another, an Ops agent checks open pull requests every morning at 9am and texts a summary.

Install is through Homebrew or an arm64 .dmg download. The app is built with Electron, React 19, Phaser 3, SQLite and the Claude Code SDK. The FAQ says the app runs locally but the models are run by Anthropic, so your code does go to the cloud. It also says Windows and Linux builds are not fully tested, and that agents sometimes refuse non-coding tasks.

How to build something similar

  • Start with two or three agents that have clear, separate roles before you add an orchestrator on top of them.
  • Give agents a way to hand work to each other, such as a shared message queue, and log every handoff so you can see where a task stalls.
  • Put an approval step in front of anything destructive, like deletes, installs and network changes, and keep it on until you trust the setup.
  • Track token spend per agent from day one. Parallel agents can run up a bill faster than a single session.

The original post

georgesayoub · View on the web

Want to build something like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week.

Join Shipping Skool — Free

More use cases

← All AI agent use cases