Foreman

What it does

Foreman lets Codex, Claude Code, or another caller hand implementation work to AI engineer workers while the caller stays architect and QA. Workers get scoped specs, isolated git worktrees, durable logs, and a collection path back into the parent repo.

It is built for work that can be split: one worker owns one bounded implementation slice, reports what changed, and leaves the final judgment to the caller.

Install

For Claude Code, install Foreman from the Eidos Marketplace:

claude plugins marketplace add eidos-agi/eidos-marketplace
claude plugins install foreman@eidos-marketplace

The Python CLI package is prepared for PyPI as eidos-foreman. Until the first PyPI release is tagged, use the plugin install above or install from the source repo.

pip install git+https://github.com/eidos-agi/foreman.git
foreman --help
foreman-mcp

Use

foreman delegate --repo /path/to/repo --engine claude "Implement the focused change and run tests."
foreman list
foreman status <worker_id>
foreman tail <worker_id>
foreman collect <worker_id>

Plugin tools

The plugin exposes delegation, status, tail, collection, notes, interruption, and control-plane tools. The important behavior is fire-and-return: delegation gives you a durable worker id, worktree path, log path, and reattach commands instead of keeping a model request open until the worker finishes.

Control plane

Foreman has a portable SQLite-backed control-plane path for local or distributed workers. Jobs, leases, events, worker ids, and status live in a generic coordination layer so future Slack, web UI, Railway, or rented-Mac clients do not need to know Foreman's worktree internals.

Boundaries

• Workers implement: they own scoped branches and tight file ownership.
• The caller reviews: tests, integration, merge, PR, discard, or send-back decisions stay with the architect.
• The daemon is explicit: live monitoring uses a small local daemon, but plugin install does not start it automatically.

See also the Eidos Plugin Store and Eidos Marketplace.