Say what you need. An AI team does the work and asks when it needs you.

1. You say what you need

   In plain words, in the AI assistant you already use, in one folder.

2. The lead hands it on

   It writes a short brief for the worker and passes questions back to you.

3. The worker does it

   With a sandboxed setup each bot works in its own workspace, separate from your project and home folders.

4. Review checks, you decide

   A second bot checks the result before it comes to you. What happens with it is your call.

• Each bot in its own workspace

  Choose a sandboxed setup to give each bot its own workspace, separate from your project and home folders.

• A check before review

  The gate command you configure, your tests for example, runs before review without an AI call. This repository's own gate adds a complexity check (scripts/complexity.py) that you can copy into yours.

• Side by side, checked first

  For a batch, WOUF checks that the tasks' declared file areas do not overlap before starting them together.

• Shared between teams

  Teams can exchange signed handoffs through a shared Git repository, without a dedicated WOUF server.

• Work you can follow

  Team events and readable summaries can live in your Git repository, so you can follow the work over time.

1. Have an AI assistant for the terminal

   For example Claude Code, Codex CLI or Antigravity. Install one and sign in once.

   • Claude Code
   • Codex CLI
   • Antigravity (agy)
   • Grok Build
   • any CLI that reads AGENTS.md

2. Make a folder and open it in the terminal

   Then start your assistant there.

   Never opened a terminal? Here is how.

   On a Mac: make the folder in Finder. Open the Terminal app, type cd and a space, drag the folder into the window and press Return.

   On Windows 11: open the folder in File Explorer, right-click an empty spot in it and choose “Open in Terminal”.

   Now type the name of your assistant, for example claude, and press Enter.

3. Paste one sentence

   The assistant sets everything up in that folder and asks a few questions, each with a recommendation. Typing your key is the one step it leaves to you.

• Host mode has no sandbox

  Docker, Podman and bubblewrap runs give a bot a separate workspace. Host mode runs without a sandbox and can reach your computer with your permissions.

• What you give a bot, it has

  Inputs, repository content and keys you hand to a bot can still reach it. A sandbox is not a guarantee against every exploit or leak.

• The network is open by default

  The shipped container setting gives open network access. Cloud models need the internet, so what a bot works on goes to the model provider.

• Rules are instructions

  Roles and questions are instructions to the models, not a guarantee that they comply.

• Legal rules are yours to check

  Whether a setup meets the law where you work depends on the model and how you run it. WOUF does not settle that for you.

All limits in the docs

These are records of the development repository, not proof that every change was made by bots. The public repository starts each release from a fresh history, so its commits do not carry them; the changelog says what came in, and docs/examples/T-0001.md follows one task from brief to approval.