ORCH — CLI runtime that orchestrates Claude Code as a typed task queue with state machine

[oxgeneral]

Hey Anthropic SDK community! 👋

I built ORCH — an open-source CLI runtime that manages Claude Code instances as a structured engineering team.

What it does

If you use the Anthropic SDK to build agents, ORCH handles the orchestration layer:

npm install -g @oxgeneral/orch
orch init

# Add Claude as an agent
orch agent add "Claude Dev" --adapter claude

# Create and dispatch tasks — ORCH handles assignment
orch task add "Implement OAuth flow" -d "Use JWT, add to src/auth/" -p 1

# Watch live in TUI dashboard
orch tui

Architecture

• State machine: todo → in_progress → review → done (mandatory review step, prevents skipping)
• Auto-retry: configurable retry on failure with context propagation
• Inter-agent messaging: orch msg send <agent-id> "..."
• Goal system: high-level goals decomposed into subtasks automatically
• Teams: group agents into specialized teams (backend, frontend, QA)
• Zero vendor lock-in: works with Claude, Codex, Cursor, or any shell command

Why I built this

I was spinning up multiple Claude Code sessions manually and losing track of task state, context, and which agent was doing what. ORCH solves exactly that — it is the "project manager" layer between you and your AI agents.

📦 npm: @oxgeneral/orch
🔗 GitHub: https://github.com/oxgeneral/ORCH
✅ 1493 passing tests, TypeScript strict mode

Would love feedback from anyone building multi-agent systems with the Anthropic SDK! What orchestration patterns are working for you?