MaxMiksa/Auto-Company: An auto-company works for 24/7 on your own PC - Windows/Linux/macOS.

An auto-company works for 24/7 on your own PC - Windows/Linux/macOS. - MaxMiksa/Auto-Company

Summary

Auto-Company is a self-hosted Python framework which runs a multi-agent AI workforce in a continuous loop on a local machine, using Claude Code (default) or Codex CLI to research, decide, and code within human-configured goals, permissions, and budgets. It ships with 14 named expert personas (for example a "CTO" modeled on Werner Vogels and a "QA" modeled on James Bach), 30+ reusable skills, and a local Product Center dashboard at http://127.0.0.1:8810/center. The project is MIT-licensed and targets macOS and Windows/WSL.

What it is useful for

It is useful when an engineer wants a persistent team of agents running 24/7 on their own PC to take a product idea from brainstorming through validation and into implementation, rather than running ad-hoc chat sessions. The framework supports up to four parallel projects by default and includes 22 independently runnable examples in examples/, such as ScopeFence for scope/price/schedule decisions, an Image Checklist tool, and a COI Chase Desk for follow-up tracking. Each cycle reads the prior state from memories/consensus.md, picks work, executes, then sleeps before the next cycle, functioning as a long-running autonomous background worker.

Documented limitations apply. Continuous operation depends on model quota and service availability; sub-agent creation and concurrency depend on the executing model and engine, so the team is not literally 14 permanent workers. Deployment, publishing, and marketing only happen when the host has the tools and authorizations to do them, and the cycle prompts are guidance rather than an enforced state machine, so a third cycle does not guarantee a deployed product.

How engineers can use it

Prerequisites are Claude Code or Codex CLI authenticated, Python 3.10+, Git, and make, on macOS or Windows 10/11 with WSL2 (Ubuntu). On macOS or inside WSL, clone the repo and run make start for a foreground loop, or ENGINE=codex make start to switch engines; make install registers a launchd (macOS) or systemd --user (WSL) daemon that auto-restarts on crash. On Windows PowerShell, .\scripts\windows\start-win.ps1 starts the daemon with -Engine codex available; companion scripts cover status, stop, monitor, last-cycle output, and the dashboard. To redirect work, stop the loop, edit "Next Action" in memories/consensus.md, then resume. The README warns that edits during a live cycle are not guaranteed to take effect immediately. For full configuration details, see i18n/en/docs/windows-setup.md and CLAUDE.md.