baro
One sentence in. A pull request out.
baro turns a goal into a coordinated team of coding agents — planned, parallelized, reviewed and merged in your terminal, on the subscription you already pay for.
A team, not a task queue.
Most parallel agents work blind — same repo, zero awareness of each other. baro's agents share what they learn while they work: what one discovers changes what the others do.
Findings ride the event bus mid-run — measured facts, not vibes. The next agent starts smarter than the last.
Failures, merges and blockers stream back into planning. The plan you end with is not the one you started with — on purpose.
Your tests are the spec. Every story faces a critic; every merged result faces the full suite before the PR opens.
You don't pick the model. baro does.
Not every task deserves the same model. Planning is where quality is won — so baro spends a frontier model there. Building is a commodity — so baro runs it cheap. Frontier-quality results at a fraction of the cost, and you never think about models.
The hard thinking — turning your one sentence into the right plan. baro routes it to a top-tier model, where paying for quality actually pays off.
Writing the code once the plan is clear — commodity work. baro runs it on a fast, inexpensive model, and escalates only the parts that need a second look.
As big as the job.
A quick fix and a repo-wide feature aren't the same work. baro reads your goal, proposes how to run it — with its reasoning — and you confirm.
A small change. One agent, in and out.
Steps that build on each other, in order.
Independent stories side by side — isolated worktrees, merges serialized and verified.
Accept the proposal, or force one with --mode focused|sequential|parallel.
Point it at a real goal.
The install takes a minute. The first pull request takes one sentence. MIT-licensed, runs on Claude Code, Codex or any OpenAI-compatible endpoint.