Wardroom is an open-source, local-first orchestrator for disciplined AI development. It runs two Claude agent roles over your git repository: a Product Manager that owns scope and canonical documents, and an Implementer that writes code and tests. You keep command, so work proceeds autonomously in bounded, test-gated iterations, and only critical actions (pushing, deploying, changing scope, destructive operations) stop and wait for your approval.
The name: a wardroom is the room on a warship where the officers plan and report. The crew works there. The captain, you, keeps command.
Pre-alpha, under active development. Nothing runnable exists yet. This repository currently holds the project's public skeleton while the core is being built. Watch the repo if the idea interests you.
Autonomous coding agents fail in two directions. "Issue in, PR out" agents produce large, unreviewable changes and lose their state to context limits. Heavyweight multi-agent frameworks bury project state in databases you can't read or fix. Wardroom takes a third path:
- All durable state lives in repository files. Plain, versionable markdown, with no hidden databases and no session memory to lose. Kill the process at any moment; it resumes from the repo.
- Documents before code. Scope, requirements, and decisions are written and human-approved before implementation runs.
- Bounded iterations that must close green. Full test suite, reversed order, lint and types, or the iteration doesn't close.
- Human command, not human babysitting. You approve gates, not every step.
- Single core with pluggable control surfaces: CLI first; Telegram bot and MCP server later, so you can supervise from a phone or from Claude chat.
- Local-first: your machine or your own server, your own Anthropic credentials. No central service, no telemetry, no account.
- Built on the Claude Agent SDK. Claude-only by design.
MIT. See LICENSE.