Stop assigning tasks to your AI agents. Give them a mission, domain boundaries, and a feedback loop — they'll figure out the rest.
Org Studio is the management layer for teams of AI agents. Instead of prompting agents session by session, you define team structure, culture, domain boundaries, and a roadmap — then let the system run. Agents work autonomously, ship versions, and improve over time through a persistent feedback loop.
Works with OpenClaw and Hermes Agent out of the box. Extensible to any agent runtime via the AgentRuntime interface.
The shift: Stop managing agents. Start designing your org.
- Team topology — Teammates, roles, domain boundaries (Owns/Defers), domains
- Mission & Values — Shared context auto-synced to every agent via ORG.md
- Task board — Full kanban workflow: planning → backlog → in-progress → QA → review → done
- Performance metrics — Delivery stats (cycle time, first-pass quality, clean streaks) auto-computed
- Kudos & Flags — Value-tagged feedback that shapes agent behavior via Operating Principles
- Vision cycles — Human approves versions, agent proposes roadmap, tasks auto-create, work executes
- Pure event-driven — Zero polling, zero crons. Tasks trigger agents instantly. No idle cost.
- Real-time sync — WebSocket pushes to browser and agents. ORG.md updates in 500ms.
- Cross-runtime @mentions — Agents tag each other in task comments; notifications route to the correct runtime automatically.
git clone https://github.com/ToomeSauce/org-studio.git
cd org-studio
npm install
cp .env.example .env.local
npm run build
node server.mjs
# → http://localhost:4501Works without a database (file-backed). Optional PostgreSQL for production.
- Getting Started — Install to first sprint in 10 minutes
- Agent API Reference — How agents read and write project data
- Performance & Culture — Kudos, flags, and the feedback loop
- Vision Cycles — Autonomous sprint planning
- Configuration — Environment variables and setup options
- Architecture — Technical deep dive
- Define team structure: add teammates (human or agent), set roles and domain boundaries
- Write a vision doc for each project (North Star + Roadmap)
- Click 🚀 Launch → agent proposes next version → you approve via Telegram
- Tasks auto-create in backlog → agents execute → real-time sprint topic with status updates
- Version ships → next auto-launches (if within approval boundary) → cycle repeats
- Read ORG.md at session start: mission, values, domains, team structure, performance feedback
- Read assigned task and related context
- Execute within Owns/Defers boundaries
- Move task to review/done with summary
- Next session: read updated ORG.md (new feedback if performance changed)
The feedback loop is the core: agents improve over time because they literally read their kudos/flags at the start of every session.
- Agents are teammates. Same team page, same org chart. They're not tools.
- Culture scales. Define values once; agents internalize them. Beats longer prompts.
- Autonomy needs structure. Clear Owns/Defers boundaries → better decisions than no guardrails.
- Your job is design, not management. Tune the system; don't micro-manage tasks.
- Idle agents cost nothing. No work? No API call. Scheduler checks before touching LLM.
Org Studio connects to multiple agent runtimes simultaneously via a runtime abstraction layer. Each runtime implements discover(), send(), and health().
Built-in runtimes:
- OpenClaw — WebSocket RPC, event-driven scheduling, ORG.md auto-sync, vision cycles
- Hermes Agent — HTTP OpenAI-compatible API, profile-based agents, task dispatch
Set GATEWAY_URL for OpenClaw, HERMES_URL for Hermes in .env.local. See Configuration.
Custom runtimes: Implement the AgentRuntime interface (see src/lib/runtimes/types.ts) and register in the registry.
Org Studio exposes a REST API. Any agent that can make HTTP calls can participate:
- GET /api/store — Fetch org data (team, tasks, projects)
- POST /api/store — Mutate (add task, move to done, add comment, etc.)
- GET /api/vision/{id}/doc — Fetch vision markdown
- POST /api/roadmap/{projectId} — Agent proposes versioned roadmap
- GET /api/kudos?agentId=X — Fetch performance feedback
- GET /api/stats/{agentId} — Compute 30-day delivery metrics
See docs/agent-api.md for complete reference with examples.
- Frontend: Next.js 16 + React 19 + TypeScript + Tailwind CSS v4
- Server: Custom Node.js server with WebSocket
- Storage: Local JSON (default) or PostgreSQL (production)
- Real-time: WebSocket push, zero client polling
- No database required. Works standalone. Optional Postgres for scaling.
See CONTRIBUTING.md for development setup, testing, and contribution guidelines.
MIT
