ORCH — open-source CLI orchestrator for multi-agent teams (TypeScript, Claude/Codex/Cursor)

[oxgeneral]

Hey AutoGen community! 👋

I've been working on a problem complementary to AutoGen — how to manage and coordinate existing CLI-based AI agents (Claude Code, Codex CLI, Cursor, OpenCode) as a typed task queue with a validated state machine.

Built ORCH — an open-source CLI runtime for multi-agent coordination:

npm install -g @oxgeneral/orch

# Register agents with different adapters
orch agent add "architect" --adapter claude --role "Senior TypeScript developer"
orch agent add "coder" --adapter codex --role "Unit test writer"
orch agent add "refactor" --adapter opencode

# Create typed tasks with priorities and scopes
orch task add "Implement auth module" -p 1 --assignee architect --scope "src/auth/**"
orch task add "Write tests for auth" -p 2 --assignee coder --depends-on <task-id>

# One dashboard
orch tui

How it differs from AutoGen:
AutoGen = framework to build agents in Python
ORCH = runtime to orchestrate existing CLI tools (Claude Code, Codex, Cursor, shell scripts)

Core mechanics:

• State machine: todo → in_progress → review → done with auto-retry on stalls/failures
• Scope locking — prevents concurrent agents from editing the same file
• Inter-agent messaging: orch msg send <id> "context" — agents share findings mid-run
• Goal system — break big goals into agent tasks automatically
• TUI dashboard — real-time visibility across all agents
• Zero infrastructure — state in .orchestry/ YAML/JSONL files, no database

For AutoGen users: ORCH can complement Python-based agents by providing the orchestration shell layer. You could run AutoGen agents via the shell adapter and coordinate them with ORCH's task queue and state machine.

GitHub: https://github.com/oxgeneral/ORCH
npm: npm install -g @oxgeneral/orch

1,657 tests · TypeScript strict · MIT · Node 20+

Would love feedback from the AutoGen community — especially on the adapter design and how it might bridge with Python-based agent frameworks!