ORCH

What It Does

ORCH is an open-source (MIT) CLI orchestrator that coordinates multiple AI coding agents — Claude Code, OpenAI Codex, Cursor, OpenCode, or any shell command — working in parallel on the same codebase. Each agent operates in its own isolated git worktree and branch, preventing merge conflicts. A typed state machine (todo → in_progress → review → done) governs all task transitions, and a mandatory reviewer agent acts as a gate before any code reaches the main branch.

The tool is structured around a department/org model: engineers define agents with roles, group them into teams, compose teams into departments (orgs), and can deploy pre-built templates for entire functions (engineering, security, content, data). A TUI dashboard provides real-time visibility; a daemon mode with structured JSON logging enables CI/CD integration via pm2 or systemd. A Claude Code /orch skill allows natural language task dispatch. The project launched in March 2026 and reached v1.0.22 by April 2026, with 1,694 passing tests.

Key Features

• Git worktree isolation: each agent works on a dedicated branch, with conflict-free parallel execution across multiple agents
• Typed state machine: todo → in_progress → review → done with cascade-fail (permanent failures propagate to all dependent tasks)
• Mandatory reviewer agent gate: no code reaches main without passing a configured reviewer agent
• Adapter-agnostic: supports Claude Code, Codex, Cursor, OpenCode, and any shell command (npm, Python, Semgrep, curl) as first-class agents
• Pre-built org templates: startup-mvp, security-dept, test-factory, content-agency, data-lab, sales-machine, and more
• TUI dashboard with real-time agent and task status; TUI Observer Mode for read-only monitoring of a running daemon
• Daemon mode (orch serve) with structured JSON logging, graceful shutdown, memory monitoring, and CI/CD mode (--once)
• Automatic retry with exponential backoff; zombie task detection and recycling for stalled agents
• Inter-agent messaging for coordination across concurrent workers
• Claude Code /orch skill for natural language task and team management

Use Cases

• Parallel feature development: Decompose a feature into independent subtasks, spawn multiple coding agents on separate worktrees, then coordinate via the state machine and review gate before merging.
• Automated security or QA pipelines: Use the security-dept or test-factory templates to run parallel audits or coverage improvement passes without manual coordination.
• Overnight autonomous coding sessions: Assign tasks at end of day, let daemon mode run agents overnight, review state-machine-gated PRs in the morning.
• Multi-tool agent workflows: Mix Claude Code agents for reasoning-heavy tasks with shell agents (Semgrep, npm test, Python scripts) in the same typed task queue.
• Teams evaluating multi-agent coding: A zero-infrastructure (no Kubernetes, no Docker, no database) entry point for teams exploring coordinated AI coding agents.

Adoption Level Analysis

Small teams (<20 engineers): Good fit. Node.js >=20 is the only requirement — no infrastructure overhead. The department templates provide a quick starting point without deep configuration. The mandatory review gate reduces the risk of silent agent failures reaching production. However, at 18 GitHub stars and no public production reports, early adopters are accepting significant stability risk.

Medium orgs (20-200 engineers): Potential fit for forward-leaning engineering teams. The daemon mode, structured JSON logging, and pm2/systemd integration support production deployment. The state machine and cascade-fail provide operational safety. However, no enterprise governance features (RBAC, audit trails, compliance) are documented, and the project has not been independently validated at scale.

Enterprise (200+ engineers): Does not fit today. No RBAC, no audit logging, no compliance certifications, unknown author, minimal community. The tool is too new and unproven for enterprise deployment in regulated or critical environments.

Alternatives

Evidence & Sources

• ORCH GitHub Repository — source, README, changelog, 1,694 tests
• ORCH npm Package (@oxgeneral/orch) — release history
• Addy Osmani: The Code Agent Orchestra — context on multi-agent coding patterns
• Bunnyshell: Agentic Development in 2026 — industry background on AI coding agent orchestration
• Composio Agent Orchestrator — primary architectural comparison

Notes & Caveats

• Unknown author, no organizational backing. oxgeneral is an anonymous individual GitHub account with no public profile. No company, funding, or team size is disclosed. This is high adoption risk for any production use.
• Very young project. First release was March 12, 2026. As of April 2026, the project is six weeks old. The rapid v1.0.x release cadence (22 patch releases in one month) signals active development but also instability — deadlocks, false success signals, and race conditions have all been documented and fixed in recent releases.
• No infrastructure required is a double-edged sword. Zero dependencies (no Docker, no Kubernetes, no database) lowers the adoption bar but also means there is no hardened runtime isolation between agents. Agents share the host OS, which has security implications for untrusted codebases.
• Non-engineering templates are aspirational. sales-machine, data-lab, and content-agency templates imply full-department autonomy that exceeds current LLM capability in business-critical workflows. These are demo-grade, not production-grade.
• LLM-reviewing-LLM limitation. The mandatory reviewer agent is ORCH’s key safety claim, but LLM reviewers have well-documented blind spots. Users should not rely solely on the reviewer gate for code quality assurance — human review remains necessary for anything beyond trivial changes.
• MIT license is clean. No BSL, no source-available restrictions, no CLA.