SuperAPI AI Directory
Partner with us中文AI newsFavorites
All toolsAI agentsORCH

ORCH

ORCH is an open-source (MIT-licensed) command-line orchestration tool for managing a group of AI agents working in parallel on a terminal or server. It unifies and coordinates various CLI tools such as Claude Code, Codex, Pi, and Cursor, and through mechanisms such as a state machine, isolated git worktrees, automatic retries, inter-department messaging, and mandatory code review, it enables multiple agents to divide labor and collaborate around the same project goal. With just one npm install, no database, cloud service, Docker, or GPU is required, and all state is kept in the local .orchestry/ directory. It is suitable both for independent developers using a small number of agents to advance a project in parallel, and for deploying a complete "department" through preset team templates; it can orchestrate any CLI workflow through shell adapters and supports long-term operation on a server as a headless daemon.

Open website

About this tool

Purpose

ORCH's core positioning is "one CLI to orchestrate all tools." It solves the problem that after you purchase and use AI tools such as Claude, Codex, and Cursor, you end up spending a lot of time copying context between multiple terminals, manually resolving file conflicts, and tracking the progress of each agent. ORCH lets multiple AI agents work in parallel around the same project goal, and users only need to set the goal without supervising each one individually. Its advertised scenario is: set a goal at 10 p.m., five agents break it down, implement, test, and review it, and you receive a pull request when you wake up.

Main capabilities

  • Multi-agent parallel coordination: can simultaneously coordinate Claude, Codex, Pi, Cursor, and any CLI tool.
  • Organized division of labor: the CTO agent is responsible for strategic breakdown, splitting high-level goals into concrete tasks and assigning them to the corresponding departments; the engineering department executes in parallel; the QA department automatically verifies, and failed tests can be returned to engineering with feedback; code review serves as a mandatory quality gate.
  • Inter-department communication: supports direct messages, team broadcasts, and a shared context store, for example notifying QA to start testing after the backend is complete, without manual routing.
  • Isolation and safety: each agent works in an independent git worktree (independent branch), and the main branch is unaffected until a merge is explicitly approved; the state machine includes a mandatory review step; it has scope overlap detection to prevent conflicts.
  • Fault tolerance mechanisms: failed tasks are automatically retried with exponential backoff; stalled tasks are terminated by zombie detection and re-queued; the state machine validates every state transition to avoid task loss.
  • Shell adapters: any tool that can run in a terminal can serve as an agent, such as npm test, python bot.py, Semgrep, curl, CRM scripts, data pipelines, etc., and gains status tracking, retries, and coordination capabilities.
  • Adapter ecosystem: provides eight adapters—Claude Code, OpenCode (Gemini, DeepSeek, any OpenRouter model), Codex, Pi, Cursor, Grok, Antigravity, and Shell. Uses the user's own API keys and tools.

Usage

Installation: a single npm install command completes it (the material does not provide the full original installation command, only stating that it is one npm install). After installation, ORCH automatically initializes andOpen the TUI dashboard, where you can add agents, set goals, and run.

Claude Code integration: after installation, the /orch skill is automatically provided in Claude Code. Type /orch and describe your requirements in natural language, and Claude will convert them into the corresponding orch commands (create agents, tasks, and goals and run orchestration), with no need to memorize CLI parameters.

You can also deploy preset teams with one click (such as the content-agency, data-lab, sales-machine, and other templates).

CLI aliases: orchestry, orch, ao.

Headless daemon and CI/CD: run 24/7 on a server via orch serve, with no TUI, outputting structured JSON logs that can be integrated with Datadog, Grafana Loki, or jq. It supports watch mode (the default, where the tick loop runs indefinitely and tasks can be added from another terminal with orch task add) and once mode (--once, which exits after processing all existing todo tasks and does not perform autonomous task seeding). It provides lock protection (only one orchestrator per project, reusing .orchestry/orchestry.lock; a second orch serve will report an error and exit), graceful shutdown (after SIGINT/SIGTERM, stop accepting new tasks, wait for running agents, save state, and release the lock), heap memory monitoring (record heap_mb on each tick), and idle throttling (record about once every 6 idle ticks). It can be deployed with pm2 or systemd.

Use as a library: ORCH is engine-first and CLI-second; the core has zero dependencies on the CLI/TUI layer, so you can import @oxgeneral/orch to build your own interface.

Suitable scenarios

  • Independent developers or small teams want to complete work that originally required an entire team in a "one person plus several agents" way.
  • Need to let multiple AI agents develop in parallel on the same codebase without overwriting each other.
  • Need to automate processes rather than only coding, such as editing, sales, analysis, security, DevOps, etc., as long as they can run in the terminal.
  • Want to run orchestration unattended on a server for a long time and integrate it with a log aggregation system.

Limitations and resource usage

  • System requirements: minimum 1-2 agents; for a complete deployment 4-6 agents are recommended; no database, cloud service, Docker, or GPU is required (LLMs are called via API and do not run locally).
  • Memory usage: ORCH itself is about 120 MB, plus about 300 MB per concurrent agent (each is an independent Node.js/Python runtime). The formula is 120 MB + N × ~300 MB: about 0.7 GB for 2 agents, about 1.3 GB for 4, and about 2 GB for 6.
  • State storage: all in the .orchestry/ directory, using pure YAML, JSON, and JSONL files, with no account registration required, and data does not leave the local machine.
  • License and costs: open-source MIT license; you only pay for the AI APIs used (Claude, Codex, etc.), and the TUI displays the token cost of each agent run in real time (example: about $4.20 for 5 agents and 6 tasks).

Information not yet provided

The materials do not give the complete original installation command, the detailed parameter tables for each CLI subcommand, the complete list and contents of the preset team templates, specific API pricing, or a runnable demo environment.For a full deployment, 4-6 agents are recommended; no database, cloud service, Docker, or GPU is required (LLMs are called via API and do not run locally).

  • Memory usage: ORCH itself is about 120 MB, plus about 300 MB per concurrent agent (each is an independent Node.js/Python runtime). The formula is 120 MB + N × ~300 MB: about 0.7 GB for 2 agents, about 1.3 GB for 4, and about 2 GB for 6.
  • State storage: all in the .orchestry/ directory, using pure YAML, JSON, and JSONL files, with no account registration required, and data does not leave the local machine.
  • License and costs: open-source MIT license; you only pay for the AI APIs used (Claude, Codex, etc.), and the TUI displays the token cost of each agent run in real time (example: about $4.20 for 5 agents and 6 tasks).

Information not yet provided

The materials do not give the complete original installation command, the detailed parameter tables for each CLI subcommand, the complete list and contents of the preset team templates, specific API pricing, or a runnable demo environment.

Related tools

Browse category