ORCH — CLI runtime that turns Claude Code into a coordinated AI engineering team

[oxgeneral]

Turn Claude, Cursor, and Codex into one team.

Built ORCH — an open-source CLI runtime that orchestrates Claude Code (and other AI coding agents) as a typed engineering team with formal state machine, auto-retry, and inter-agent messaging.

If you are using Claude Code, ORCH adds the coordination layer:

• Run multiple Claude Code instances in parallel, each with a specialized role
• State machine enforces todo -> in_progress -> review -> done lifecycle
• One Claude instance can message another: orch msg send agt_reviewer check this
• Shared context: orch context set codebase-overview lets all agents read shared state
• TUI dashboard shows all Claude Code agents running in real time

Architecture overview:
npm install -g @oxgeneral/orch
orch agent add backend --adapter claude --role Senior Backend Developer
orch agent add reviewer --adapter claude --role Code Reviewer
orch task add Implement JWT auth --assignee backend
orch tui # live view of all Claude Code agents

What makes ORCH different from just running multiple Claude Code sessions:

• Tasks cannot silently complete — mandatory review gate before done
• Failed Claude Code runs trigger retrying state and automatic restart
• Agents share context without copy-pasting between terminals
• Goal hierarchy: high-level goals decompose into tasks assigned to specific Claude agents

1493 tests, TypeScript strict, MIT license. 5 adapters: claude, opencode, codex, cursor, shell.

GitHub: https://github.com/oxgeneral/ORCH

Curious what multi-agent Claude Code workflows people here are building?

[affaan-m]

This is very aligned with where serious Claude Code workflows end up: multiple sessions, explicit roles, shared context, and review gates.

The parts I would evaluate hardest are the boundaries:

1. What context is shared between agents and what stays private?
2. How are secrets and local paths kept out of cross-agent messages?
3. Can a failed agent retry without repeating destructive actions?
4. Is the review gate policy-based or just another model saying yes?
5. Can users inspect the state machine and message log after the run?

ECC has been moving toward similar orchestration patterns through skills, worktrees, tmux/dmux-style flows, and verification gates. ORCH looks like a useful independent runtime to compare against that direction.