Open source · MIT
agent-team
A Claude Code plugin that seats a team of AI roles in a project: a CEO, an Engineering lead, a Bar Raiser and whichever other seats the repo calls for, each with its own rulebook, its own workers on the right model, and a clear line on what it decides. A re-evaluation measures how each seat actually works and proposes changes.
MITOpen source
What I built
- A team sized from evidence. The scan reads manifests, file names, CI, deploy configs and git history, and each proposed seat cites what raised it.
- A measured re-evaluation loop. It reports each seat's activity over the last 30 days and what changed in the repo, then proposes changes. What the owner declines is remembered.
- Seats that check each other. Work goes to a Bar Raiser for review, and seats exchange ASK, HANDOFF and RESULT messages instead of free-form chat.
- Model tiers per seat. Typing goes to cheap, fast models and judgement to strong ones, steered from above.
- Checked on every push. GitHub Actions runs the rulebook checks and seats a fresh project with the scaffold on Python 3.9 and 3.12; unit tests run on repo fixtures.
- Runs my own builds. It seats the teams behind my private products, so every rule is tested on real work first.
Technical highlights
- 01
Python standard library only: no dependencies to install.
- 02
Releases follow semantic versioning, with every release listed in CHANGELOG.md.
- 03
Installs as a Claude Code plugin from its own marketplace, or by linking the skills from a clone.
- 04
Runs on Windows as well as macOS and Linux, since 1.3.1.
By the numbers.
- 3.9 + 3.12
- Python versions CI runs on
- 3 teams
- Offered per repo: lean, recommended, full