Agent Stack

agent-stack

Tell it what to build. Get tested, reviewed, merged software back.

agent-stack turns one Linux computer into a software team of AI coding agents. You talk to one agent, the operator:
"onboard this repo, here is what I want". It sets up a team for the project, the team plans the work, writes the tests
first, builds, uses the result the way a person would, has a second AI family review it, and merges it through a gate.
You approve the plan and the risky steps. Everything else runs on its own.

It is glue and configuration around existing tools: OpenRig runs the
teams, Claude Code and Codex CLI are the agents' workspaces, CLIProxyAPI
pools your subscriptions, TypeSafe Jev makes the small typed decisions (who builds this, may
this merge), and Playwright gives the agents a real browser.

Watch the fleet: rig-console

rig-console is a full-screen, read-only console for every team on the machine. It opens with what needs you,
then lets you follow a project through its agents, tasks, reviews and merge gate. Capacity shows recorded usage
and reset times. Missing measurements stay unknown; cached fleet data is marked stale.

Try it from this checkout with node console/src/main.ts --fixture console/fixtures/v3.json. No daemon or account
is needed. The gallery, keys and examples are in rig-console in detail.

How it works

- Operator. An always-on agent in OpenRig's kernel rig. You talk to it; it sets up and watches the project teams.
- Rig. One project's team: a lead, an architect, test authors, builders, QA, reviewers of two AI families, and a
  merge owner. Four sizes: core (4 seats), small (10), build (14), full-stack (27).
- Seat. One agent with a role, its own git worktree and its own conversation, named -@ (for
  example coord-lead-claude@shop).
- Queue. Seats…