Claude Foreman

Foreman

Foreman lets Claude delegate implementation work to external coding-agent CLIs. Claude acts as the foreman: it plans the work, writes a self-contained brief, reviews the diff, and verifies the result. A worker CLI does the implementation.

Quick start

Install Foreman as a Claude Code plugin:

Or with the skills CLI:

Then run in claude :

Fill in the brief before dispatching. It should state the goal, relevant context, allowed file scope, rules, and an exact verification command. The worker does not have access to the foreman's conversation. After the run, inspect the diff and run the checks yourself; a worker's report is not verification.

How it works

By default, Foreman runs workers headlessly. Write-mode tasks run in an isolated Git worktree on a separate foreman/ branch, with the selected adapter's sandbox enabled. Read-only tasks run in the repository without edits. Workers should not commit, push, or modify the user's branch. The foreman reviews and verifies the changes before they are integrated.

The included adapters support OpenAI Codex (codex exec) and Google Antigravity (agy -p). Foreman can also support other coding-agent CLIs through adapters; see Adding a harness for the adapter contract and checklist.

Sandbox and permission behavior depends on the adapter. Codex runs with its sandbox on. Antigravity also keeps its sandbox on; it blocks writes outside the worktree and Git, but does not block network access. In headless mode, Antigravity shell commands are auto-denied unless the user has explicitly opted into --skip-permissions. Keep briefs scoped to the task and review worker changes before integration.

Optional live mode

Use --live when you want to watch or co-drive a worker in its terminal UI. This mode uses a shared tmux session; the…