Turn a repo into an AI-executed backlog
Maestro plans your git project, splits it into skill-matched tasks, and runs them through the AI coding agents your developers already use — Claude Code, Codex, Gemini, Aider, or manual. Live progress, full control, zero lock-in.
Bring your own keys. Your code never leaves your machines.
$ cd ~/work/payments-api
$ maestro
✓ detected github:acme/payments-api
✓ identity acme-work · 3 projects linked
✓ connected ws://studio.acme.dev (lead session)
→ 2 tasks dispatched to this machine
#418 Add idempotency keys to /charge [in_progress]
#421 Backfill webhook signature checks [ready]The orchestration layer between your plan and your agents
Most AI coding tools stop at a single chat window. Maestro coordinates many agents, many developers and many projects against one shared plan.
Multi-Git, no lock-in
GitHub, GitLab (self-hosted/Enterprise included) and Bitbucket behind one provider interface. OAuth, webhooks and PRs handled for you.
One CLI, many projects
A single Rust daemon holds parallel connections to every project — and to separate work and personal Studios — from one machine.
Auto-detect by git remote
cd into a repo, run maestro. It normalizes the remote, finds the project and connects with zero flags.
Multi-provider AI + BYOK
Claude Code, Codex, Gemini, Aider or manual mode. Keys belong to the developer and stay in the OS keyring.
Skill-matched dispatch
A skill matcher routes each task to the right person or agent by skill slug and 1–5 proficiency.
Persistent project knowledge
maestro analyze ingests architecture and reusable code so agents reuse what exists instead of rewriting it.
Plan → dispatch → execute → review
Four moving parts, one shared state machine. Each step is observable in Studio in real time.
-
1
Plan with an agent
Describe the goal in Studio. A planning agent drafts a phased plan; you edit it in the plan editor and activate it.
-
2
Skill-matched dispatch
The dispatcher scores every task against developer and agent skills, then creates runs picked up automatically by connected machines.
-
3
Agents execute
The CLI runs each task through the chosen provider on the developer’s machine, opening branches and PRs as it goes.
-
4
Review & loop
Reviewers comment; a verdict either merges or spawns rework subtasks — the loop continues until approved.
Plan with an agent
Describe the goal in Studio. A planning agent drafts a phased plan; you edit it in the plan editor and activate it.
Skill-matched dispatch
The dispatcher scores every task against developer and agent skills, then creates runs picked up automatically by connected machines.
Agents execute
The CLI runs each task through the chosen provider on the developer’s machine, opening branches and PRs as it goes.
Review & loop
Reviewers comment; a verdict either merges or spawns rework subtasks — the loop continues until approved.
Stop paying to rewrite code you already have
Every analysis run catalogs your functions, services, components and conventions with vector embeddings. Before an agent writes anything, it sees what already exists — fewer tokens, fewer duplicate implementations, more consistent code.
- Typed artifacts: functions, helpers, services, components, modules, patterns
- Versioned context files snapshot architecture and workflow on every change
- pgvector search scoped to your project or shared across the workspace
- Agents query the catalog first, so they extend instead of reinventing
$ maestro analyze
→ scanning github:acme/payments-api
✓ RepoContext framework=laravel stack=php,postgres
✓ ContextFiles codebase v3 · workflow v2 (snapshotted)
✓ KnowledgeArtifacts 142 indexed (pgvector embeddings)
services 38 components 21
helpers 49 patterns 11 modules 23
→ agents will now reuse these before writing new codePoint Maestro at your repo today
Free to start. Bring your own AI keys. Self-host the Studio or run it locally.