Open source · v0.3.1

Orchestrate AI agent teams from your terminal

Run a coordinated team of AI agents in parallel — with git worktree isolation, inter-agent messaging, and state machine governance. No cloud. No config files. One npm install.

15 ready-made agents 4 AI providers MIT license
zsh — my-project
$ orch run --all --watch
  ● Running — 3 agents dispatched
 
  backend-a implementing auth module [feature/auth]
  qa-a writing test suite [test/auth]
  reviewer awaiting review tasks
 
  ✉ backend-a qa-a "Auth module ready for testing"
  ✓ tsk_a1b — review complete — merging [feature/auth] → main
Scroll
v0.3.1 — MIT
N+
Parallel Agents
15
Ready-Made Agents
1
Command to Start
0
Cloud Dependencies

Plug into the tools you already use

Claude
Anthropic
native
Codex
OpenAI
native
Cursor
Cursor CLI
native
Shell
Any CLI Tool
universal
MIT licensed 90 source files 100% TypeScript Zero runtime deps

The coordination problem

One agent is a toy.
A team is a force

Without orchestration
With ORCH

Manual juggling

You copy-paste output between AI tools manually. Context gets lost. Mistakes compound.

productivity killer

Automatic context flow

Agents message each other directly. Context flows through shared store and prompt injection.

coordinated

Agents in silos

Parallel agents overwrite each other's files. Merge conflicts pile up. Work gets lost.

wasted effort

Worktree isolation

Each agent gets its own git worktree. Parallel execution without file conflicts. Auto merge-back.

safe parallelism

Zero governance

No review process. No state tracking. A failed run leaves broken state with no way back.

high risk

State machine governance

Every task flows through todoin_progressreviewdone. Nothing gets lost.

auditable

Cloud lock-in

Most tools require cloud accounts and send your code to external servers. No control over your data.

vendor dependency

Zero cloud, fully local

Everything runs on your machine. State stored in .orchestry/ as YAML, JSON, JSONL. Git clone and run.

private

Constant babysitting

You wait for each AI run to finish. No parallel execution. No automatic retries when things fail.

time sink

Autonomous execution

The orchestrator dispatches, monitors, retries with backoff, and merges — unattended. You review the results.

automated
Stop babysitting agents. Start orchestrating them →

See for yourself

Five commands to full orchestration

zsh — my-project
$ npm install -g @oxgeneral/orch
  ✓ orchestry installed
$ cd ~/my-project && orch
  ✓ initialized · Created .orchestry/
  → Launching TUI dashboard…
$ orch task add "Implement auth module" -p 1
  ✓ Created tsk_a1b — priority 1
$ orch run --all --watch
  ● Running — 3 agents active · scope isolation · auto merge-back
    backend-a: implementing auth module [feature/auth]
    qa-a: writing test suite [test/auth]
    reviewer: waiting for review tasks
$ orch task review tsk_a1b --approve
  ✓ Task approved · merging [feature/auth] → main

See it in action

Real-time TUI dashboard with task management, agent activity feed, team coordination, and inter-agent messaging.

TUI Dashboard

Ready to orchestrate? One command to install. One command to start.

Get Started

Built for
precision

12 features · 4 categories
Coordination new

Team Intelligence

Agents that communicate, coordinate, and build on each other's work — like a real engineering team.

  • Agent Teams — group agents under a lead, broadcast context, coordinate work
  • Inter-agent Messaging — direct messages, broadcasts, injected into prompts at dispatch
  • Shared Context — key-value store readable by all agents, LiquidJS templates
Intelligence new

Autonomous Execution

Set a high-level goal. Agents decompose it into tasks and execute without hand-holding.

  • Goals & Autonomy — define goals, agents generate and execute tasks autonomously
  • Reactive Dispatch — sub-second task pickup, no polling, events trigger agents immediately
  • Smart Retries — exponential backoff, stall detection, zombie cleanup
Safety

Bulletproof Execution

Every transition validated. Every agent isolated. Every change reviewed before merge.

  • State Machinetodoin_progressreviewdone
  • Worktree Isolation — each agent gets its own git worktree, parallel without conflicts
  • Auto Merge-back — agent finishes, changes merge to main, no manual git juggling
Developer Experience

Your Terminal, Your Rules

Full-screen TUI, rework loops, and zero infrastructure. Clone and run.

  • TUI Dashboard — live tasks, agent activity, token usage, keyboard-driven
  • Rework Loop — reject with feedback, the agent retries with your notes
  • Zero Infrastructure — no cloud, no DB, YAML + JSON in .orchestry/

Only in ORCH

Agent collaboration, not just execution

Stop running agents in isolation. Build coordinated teams where agents communicate, share discoveries, and delegate subtasks — minus the standups.

CTO LEAD Backend QA Reviewer
  • Team leads — assign a lead agent who routes work and resolves conflicts
  • Direct messaging — agents send targeted instructions to specific teammates
  • Broadcasts — push announcements to the full team or scoped groups
  • Mailbox delivery — messages land in agent prompts at dispatch time
  • Shared context store — key-value pairs any agent can read or write
  • Automatic work distribution — teams claim tasks from a shared pool
Agent Teams CLI
# Create a team with a lead
$ orch team create platform --lead architect
  ✓ Created team "platform" → team_k2m

# Add a member
$ orch team join team_k2m backend
  ✓ Agent backend joined team platform

# Direct message
$ orch msg send backend "Use PostgreSQL for the new schema"
  ✓ Message sent → backend

# Broadcast to team
$ orch msg broadcast "API v2 spec is ready" --team team_k2m
  ✓ Broadcast sent to 3 agent(s)

# Share context
$ orch context set db_schema "users(id,email,role)"
  → Shared with all agents

How it works

Four steps to orchestration

01

Install

One package, no dependencies. Ready in seconds.

npm i -g @oxgeneral/orch
02

Define work

Add tasks with scopes, priorities, and agent assignments — or let ORCH auto-assign.

orch task add
03

Run

Agents execute in parallel on isolated worktrees, messaging and sharing context.

orch run --all --watch
04

Review & ship

Approve, reject with feedback, or let auto merge-back close the loop.

orch task review --approve

Built for every workflow

What teams build with ORCH

Startup

Weekend MVP Sprint

Define your vision as a goal. ORCH spins up backend, frontend, QA, and reviewer agents that build your API, UI, tests, and landing page in parallel. Ship a tested MVP in 48 hours, not 3 weeks.

Agents: Backend A/B, Front-End, QA, Reviewer, Marketer
Engineering Team

Sprint Backlog Blitz

Load 18 sprint tasks with dependencies. ORCH dispatches agents across isolated worktrees, respects ordering, retries failures, and auto-merges. Engineers wake up to draft PRs, not empty boards.

Agents: Backend A/B, Front-End, QA A/B, Reviewer
Migration

JS → TypeScript at Scale

Agents convert modules in parallel, each in its own worktree. QA runs tsc --noEmit after each merge. Reviewer rejects any any usage. Main branch stays green at every step.

Agents: CTO, Backend A/B, QA, Reviewer
DevOps

Automated PR Review Pipeline

Four agents review every PR in parallel: security scanning, performance analysis, style enforcement, and test coverage. A CTO agent synthesizes a single merge verdict in under 10 minutes.

Agents: Security, Performance, Style, QA, CTO
Creative

Product Launch War Room

Ship code and content simultaneously. Engineering team closes features while Marketing team writes blog posts, social threads, and docs — all from the same goal, with agents sharing context across teams.

Agents: Backend, Front-End, Content Creator, Marketer, Growth Hacker
Security

Multi-Layer Security Scanning

Chain SAST, dependency audit, and secret detection into one pipeline. Agents correlate findings across layers, deduplicate, assign severity, and auto-fix high-priority items in isolated worktrees.

Agents: Shell (Semgrep, Trivy, Gitleaks), Bug Hunter, Reviewer
Open Source

Contributor PRs at Scale

25 open PRs from first-time contributors? ORCH checks CLA, runs tests, reviews logic, and posts structured feedback within hours. You focus only on PRs that are ready for a human judgment call.

Agents: Shell (gh), QA, Reviewer, CTO
QA

Test Coverage Blitz

From 40% to 80% coverage without a dedicated sprint. Agents claim uncovered modules, generate meaningful tests in parallel, and QA rejects tautological assertions. You review, not write.

Agents: Shell (c8), Backend A/B, QA A/B, Reviewer
Data

Analytics Pipeline

Drop three CSVs, get an executive report by morning. Shell agents clean data, DuckDB joins and computes KPIs, matplotlib generates charts, and a Content Creator writes the narrative with anomaly callouts.

Agents: Shell (pandas, duckdb, matplotlib), Content Creator

Frequently asked

Questions

Is ORCH free?
Yes, completely free and open source under MIT license. No premium tiers, no usage limits, no cloud accounts needed.
What AI models does it support?
ORCH ships with 4 adapters: Claude (Anthropic), Codex (OpenAI), Cursor, and a universal Shell adapter that works with any CLI tool. If it takes a prompt and returns output, ORCH can orchestrate it.
Will agents mess up my codebase?
Each agent runs in an isolated git worktree on its own branch. Changes go through a review step before merging. You approve or reject every change.
Does my code leave my machine?
Only when communicating with AI APIs (same as using any AI tool directly). ORCH itself stores everything locally in .orchestry/ — no telemetry, no cloud state, no external dependencies.
How is this different from running Claude/Cursor directly?
ORCH coordinates multiple agents simultaneously: state machine tracking, inter-agent messaging, shared context, automatic retries, worktree isolation, and merge-back. It's the orchestration layer on top of your existing AI tools.
Can I use it with my existing project?
Yes. Run orch in any git repository. It creates a .orchestry/ directory and you're ready to go. No configuration required — sensible defaults and pre-configured agent templates included.
What happens when an agent fails?
ORCH retries with exponential backoff. If an agent stalls, the process is killed and the task is re-queued. Failed runs preserve full event logs. No state is lost.
Get started in 30 seconds

Start shipping with
your AI team

One command to install. One command to orchestrate.

$ npm install -g @oxgeneral/orch click to copy
then run
$ orch click to copy
MIT licensed 15 agents included Zero cloud deps 1 cmd to start