Introduction
baro is a CLI that takes a goal, spawns parallel AI coding agents, and opens a pull request when they're done.
baro lets you point a swarm of AI coding agents at a real task — a feature, a refactor, a sweep of fixes — and walk away. They plan together, work in parallel, review each other's diffs, and open a pull request when they're done. Built on top of Mozaik as a set of reactive participants on a shared event bus.
Get started
Installation
Install baro from npm, set up Claude Code or an OpenAI API key.
Quickstart
Run your first goal end-to-end and open a PR.
LLM providers
`--llm claude` vs `--llm openai` — what each path does, model defaults per phase, when to pick which.
Quick mode
Surgical fast path for trivial goals — typos, renames, version bumps.
Troubleshooting
Diagnostic flow for failed runs — start here when something doesn't work.
How it works
baro runs a five-phase pipeline:
- Discover — the Librarian participant indexes the codebase as Story Agents explore it.
- Decide — a one-shot Architect agent reads the repo and emits a
DecisionDocumentwith the exact file paths, schema names, API shapes, and library choices the rest of the run will use. One mind, decided up front. - Decompose — the planner takes the design spec and breaks the goal into a DAG of small, parallel-friendly stories.
- Execute — Story Agents run concurrently on a fresh feature branch, each handed the same authoritative spec. Critic and Surgeon participants watch the run and steer it when reality diverges from the plan.
- Ship — the Finalizer composes a full PR body (stories table, commit SHAs, diff stats, run summary) and opens the pull request.
Resources
baro on GitHub
Source, issues, releases.
baro-ai on npm
Published CLI. `npm install -g baro-ai`.
Engineering blog
The story of why baro exists and how it works.
License
baro is released under the MIT License by the JigJoy team.