orc

Turn one prompt into a team of AI agents that plan, build, review, and ship.

Your tools. Your workflow. Your rules. No daemon, no server, no lock-in.

orc /ork/ — a creature known for strength in numbers, brutal efficiency, and an unwavering commitment to getting the job done. Also: an orchestration layer for AI agents. Coincidence? Absolutely not.

---

The Problem

You ask an AI agent to build a feature. It works for 20 minutes, dumps a half-finished diff, and you spend the afternoon cleaning up. Try three agents at once and you get merge conflicts, duplicated work, and zero visibility into what's happening.

AI agents are powerful. But without structure — a plan, isolated workspaces, code review, and a delivery pipeline — they're chaos.

Orc gives them structure.

Key Concepts

Four ideas are all you need to understand orc:

Concept	What it is	Think of it as...
Goal	A deliverable with its own git branch. "Add rate limiting" is a goal.	A ticket on your board
Bead	A focused work item within a goal. One engineer, one worktree, one bead.	A subtask — small enough to complete and review in one pass
Orchestrator	An AI agent that coordinates work but never writes code. Three tiers: root (cross-project), project (manages goals), goal (manages beads).	A tech lead who delegates but never opens an IDE
Worktree	An isolated git checkout. Every agent — engineers, goal orchestrators, and project orchestrators — works in its own copy of the repo. No conflicts, no pollution of your main workspace.	Each agent gets their own sandbox

Everything else is configuration. Deep dive →

The 30-Second Version

You describe the work. Orc fans it out to parallel goals — each runs its own lifecycle: plan, decompose into work items, dispatch engineers in isolated worktrees, review every piece, merge, and deliver. Every phase is a configurable hook. Deep dive into the lifecycle →

A goal orchestrator (left) dispatches and monitors engineers working in isolated worktrees (right).

What Happens When You Run Orc

orc myapp
> "Add rate limiting to all public API endpoints. Use Redis for the token bucket."

Here is what happens next: