Point AntColony at a code folder and its engineering team runs Claude Code on it in a safe copy, with its terminal shown live. Every change is specced, built, checked and reviewed, and nothing lands until you press Merge.

The line

Brief to merge, the same way every time.

Watch a change ride the line. The reviewer's three lamps are the acceptance criteria: in the first round one misses, so the change goes down the return lane to the builder. In the second, all three light.

Then it waits for you. to land it on your branch, or for another round.

Nothing lands without you
  1. Brief

    What is wanted and why, who it's for, what done means, and the constraints. The step that matters most.

  2. Spec

    Claude Code reads the code without changing it and writes the goal, numbered acceptance criteria, the approach with real file paths, a test plan and risks. Open questions stop the run for you.

  3. Build

    The builder continues from the spec, with it as its contract, in a safe copy on its own branch.

  4. Checks

    Your tests and checks run. Failures go straight back to the builder.

  5. Review

    A separate, read-only reviewer judges the change against every acceptance criterion. Any miss sends its findings back for another round.

  6. You merge

    The run shows each criterion ticked, the checks and the review, and one button: Merge.

Step by step

Done means every criterion met.

The spec step is on for every build. Small changes get reviewed too. Review rounds stop at three by default, so a stuck change comes back to you instead of looping.

In the app

Calm enough for a beginner.

One window, one list of runs, one status for each.

One run, one status

Working, needs you, done, failed or stopped, a few plain words for where it is, and the one next step. The window, the colony and the ant all say the same thing.

Summary, Live, Changes

What happened and the button for the next step; Claude Code's steps and the run's terminals; the diff, checks, pictures and video.

Terminals like VS Code's

Every project has real shells. Yours are never limited. An ant's open only in a workspace, always sandboxed, and you can watch and type into them.

Permission asked once

When you add a project you say what the team may do there, what must pass, and what one task may spend.

Tied to the colony

Ants in the Engineering nest reach the team through its connection. Stopping a colony run stops the coding it started.

Set up in one step

Tell the Queen โ€œdeploy the engineering teamโ€: a Tech lead, a Builder, a Tester and a Reviewer, with their tools.

Coding agents it finds on your Mac

  • Claude Code
  • Codex
  • OpenCode
  • Gemini CLI
  • Cursor Agent
  • Aider
  • Amp
  • Goose

Questions

Which coding agents does AntColony support?

Claude Code first: it runs headless with its own sandbox, cost cap, session resume and live view. Other command-line coding agents found on your Mac, such as Codex, OpenCode, Gemini CLI, Cursor Agent, Aider, Amp and Goose, run the simpler way.

Can it change my code without asking?

No. Work happens in a safe copy on its own branch, you choose once per project what the team may do, and nothing lands until you press Merge.

Can it open pull requests?

Yes, if you connect GitHub. Pull requests are opened as you, and you can ask for them to be opened automatically in the project's settings.

How do I set it up?

Setup is one page: Homebrew installs Claude Code and the GitHub CLI in a terminal on the page, then you sign in with the buttons. Claude Code's location and sign-in are checked once and remembered.

Give the colony a bug tonight.

Download the app and watch a simulated colony work, no key needed. Add your own key when you want it working for you.

Download for Mac