Maestro

Maestro runs multi-agent workflows. You write a short JavaScript script that says which agents run, in what order, and what happens to their answers. Maestro executes the script with a plain runtime: each agent() call starts a model with a clean context, and the script combines, votes on and passes on the results. The control flow is code, so it is the same on every run. The reasoning is in the agents.

export const meta = { name: 'two-views', description: 'Ask two reviewers, then one judge' }

const reviews = await parallel([
  () => agent('Review src/journal.ts for correctness bugs.'),
  () => agent('Review src/journal.ts for missing tests.'),
])
return agent(`Merge these two reviews into one ranked list:\n${reviews.filter(Boolean).join('\n---\n')}`)

What you get

Resume. Every agent’s answer is stored. Change one prompt and only that agent, and the ones that depend on it, run again.
Limits in code. A budget in output tokens and a cap on agents, enforced by the runtime, not requested in a prompt.
Isolated writers. An agent that changes code works in its own git worktree and branch.
Two model providers. Agents run on Claude, on Codex, or on both in one run.
A timeline. Watch a run in the browser while it happens.

Start here

Quickstart

Connect Maestro to Claude Code and ask for a workflow. No code to write.

Getting started

Write a first workflow yourself, run it and resume it.

Writing workflows

The script API and its rules.

Claude Code

Let Claude write and start workflows for you.

Get started

Workflows

Using Maestro