Maestro is a local-first coordination system for work shared between humans and coding agents. It keeps work items, decisions, sessions, evidence, dispatches, and handbacks in a SQLite store inside each repository's Git root, runs on Bun, and needs no background service. Everything is driven from the maestro CLI.
It has three layers. A mechanism kernel owns the store, event log, sessions, command dispatch, plugin loading, and readiness projection, and imposes no workflow policy. Plugins add verbs and optional policy gates (such as proof and breakdown) that a repository enables or disables. Recipes and skills are Markdown working methods: install sets up ten managed skills covering bundles, design, work, verification, and read-only explore, diagnose, coach, and questionnaire engagements, and links them for Claude Code.
Work moves through OPEN, ACTIVE, RETURNED, and DONE, with dependencies, leases, notes, and layered proof (source, artifact, installed, live, journey) where a claim stops at the last proven layer. Decisions are immutable records that are drafted, locked, and superseded, and attention reports flag stalled leases, repeated failures, stale decisions, scope collisions, and unreviewed handbacks at read time.
Supervised teams (SLP) run a Hub Supervisor, a Team Supervisor, a Lead, and Peers as Claude Code or Codex sessions in a Herdr workspace, each launched from a Maestro-owned profile so its mandate is its system prompt. Roles use nine operations, the owner can mark themselves away so the Hub records provisional rulings, and a runtime pane nudges stalled seats. SLP is a cooperative protocol, not a security sandbox. Maestro also writes managed hooks and short pointer blocks for Claude Code and Codex so each session records itself and sees its current brief.
Features
- Repository-local store: work, decisions, sessions, evidence, dispatches, and handbacks kept in SQLite in the Git root, with no daemon
- Work lifecycle: work trees, dependencies, leases, notes, claims, cancellation, and a readiness view of startable and gated work
- Decision records: draft, locked, superseded, and provisional decisions with rationale and work links
- Layered proof: source, artifact, installed, live, and journey layers, with untested links named instead of rounded up
- Supervised teams (SLP): Hub Supervisor, Team Supervisor, Lead, and Peers coordinating through nine operations in a Herdr workspace
- Attention and brief: read-time detection of stalled leases, repeated failures, stale decisions, scope collisions, and unreturned dispatches across registered repositories
- Search and glossary: one search over work, decisions, notes, terms, memory facts, events, and imported records
- Hub memory: promotes facts from Claude and Codex memory buffers through supersession, dedup, and evidence gates
- Graph runs: executes the deterministic nodes of a multi-agent path from one Markdown graph and hands agent nodes back to the driving harness
- Harness integration: managed hooks and adapters for Claude Code and Codex that record sessions and print the current brief
- Observer mode: a read-only store mode where mutating commands fail and plugins are not loaded
- Recipes and plugins: Markdown working methods and removable policy plugins at built-in, global, and repository scope
