Don't let one coding agent grade its own homework.
Maestro is a local supervision layer for the coding agents you already use. Bring Claude Code, Codex, Gemini, OpenCode, or your own CLI-based agent: Maestro gives them roles, review gates, persistent artifacts, and a human approval loop. It does not replace your tools. It makes their work inspectable.
v0.43.0 alpha. Free. macOS Apple Silicon and Windows x64.
Origin
Why Maestro exists
Maestro started when I started building a web app with a coding agent.
The first thing I asked for was a roadmap. The agent produced one full of Kubernetes, Redis, and Grafana — infrastructure for thousands of customers I didn't have.
Then I started implementing. Hardcoded values were everywhere, impossible to trace which one overrode which. The project crashed three times.
Writing development principles helped a little. But I never had time to truly review the plans, and during implementation the agent often improvised away from them.
What finally worked was using one agent to create plans and another to review them, iterating until they converged. The output was substantially more trustworthy. But I was spending most of my time copy-pasting between models.
So I built Maestro to automate the loop. Then I added a third agent — the Conductor — to summarize each round and point me at the decisions I actually need to make.
That’s Maestro: a creator, one or more reviewers, and a Conductor that keeps the loop legible.
The goal is not to own the coding agent. The goal is to make agentic coding work reviewable, repeatable, and auditable.
Tool agnostic
Bring your own agents
Maestro is not tied to one model or one coding assistant. A workflow can use one tool as the creator, another as the reviewer, multiple reviewers in cross-check, and a read-only Conductor for summaries. The agents are replaceable; the supervision loop is the durable part.
- Use different tools for different roles —Claude Code for implementation, Codex for review, Gemini for critique, OpenCode for local OSS workflows, or another CLI-based agent as support lands.
- Keep the artifacts —Prompts, plans, reviews, address notes, implementation output, findings, and final reviews are written to disk in your project.
- Adopt the parts you need —Use the desktop loop end to end, or use its review and approval stages around an existing agent workflow.
How it works
From task to done, in six steps
Maestro runs six steps. Here's what each one is for — and what it prevents.
- Task —You describe the work in plain language. No template, no required format.
- Plan —A creator agent writes a complete implementation plan: what changes, in which files, in what order. You see the design before any code is written.
- Review —One or more reviewers critique the plan, with findings flagged by severity. The creator addresses, the reviewer re-reviews. Iterates until they converge. This is where a self-reviewing agent would have stopped too early.
- Approve —You read the converged plan and decide. No code is written until you approve.
- Implement —Maestro implements the plan one section at a time. After each section, the reviewer verifies the work matches the plan. Catches the agent improvising away from its own plan mid-implementation.
- Validate —A final comprehensive review across all sections, plus any manual checks only you can run.
Walkthrough
See it in action
A 60-second walkthrough of a real workflow — task to done.
Scope
What Maestro can and can't do
If you want a persistent personal agent that remembers you, schedules tasks, and lives in chat, Maestro is not that. Maestro is for supervised coding workflows where review and approval matter.
What Maestro does today
- Provides a local, auditable plan → review → address → approve → implement loop with any combination of supported coding agents (Claude, Codex, Gemini, OpenCode).
- Lets you assign different agents to different roles. Claude creator with a Codex reviewer, two reviewers in cross-check, or all-Claude — your call.
- Keeps the loop legible through the Conductor, a third agent that reads plan and review output, summarizes what each side said, and points you at the decisions to make.
- Persists every round as a file on disk — task, plan, reviews, address notes, implementation outputs. Inspectable later, diffable between rounds.
- Implements approved plans section by section, with per-section review, checkpoints you can pause at, and a final comprehensive review.
- Shows structured findings and click-through navigation from finding to source.
- Can be used as an end-to-end desktop workflow, or as a review and approval surface for teams experimenting with agentic development practices.
What Maestro can't do
- macOS is signed and notarized; Windows is not signed yet. Gatekeeper should accept the macOS build on first launch. Windows users may still see a SmartScreen warning; bypass instructions are in the quick start, and Windows code signing is on the roadmap.
- It doesn't bring its own coding agents. You install and authenticate Claude Code, Codex, Gemini, or OpenCode yourself. Maestro orchestrates them; it doesn't replace them. If a tool can run as a coding CLI, it should eventually be able to fit into the loop.
- No token/cost visibility yet. Long multi-agent loops can be expensive. Watch your provider dashboards until this lands.
- Reviewers are LLMs, not human experts. They catch a lot — especially when the reviewer is a different or stronger model than the creator — but not everything. The final judgment is yours.
- No shared cloud workspace yet. Maestro is currently local-first and single-machine. There is no cloud sync or organization admin surface in the alpha.
Download
Download Maestro
v0.43.0 alpha. Free. macOS Apple Silicon and Windows x64.
Before your first run, you'll need a supported coding agent installed. The Quick Start covers installation, Windows first-launch warning guidance, and getting your first workflow running.
Community
Community
The Maestro community is on GitHub.
Discussions for ideas, feature requests, and workflow stories. Issues for bugs. I'm building this on my own, so replies may take a bit — everything gets read.