How it works

The premise

A self-hosted tool that takes an app idea written in plain English and builds a working codebase, committed to git. The work is done by layers of agents running in parallel, each in its own git worktree.

A Conductor turns the idea into a plan. Section Leaders break their parts of the plan into tasks, and Musicians carry out those tasks in parallel. It commits as it goes, tracks the tokens and dollars spent on every task, and shows the plan, live progress, git history and costs in a desktop app. You can change all of it: how many layers there are, the prompts, models, tools, permissions and budgets.

What’s included

Three layers of agents, each working in its own git worktree

Commits as it goes, with token and cost tracking per task

Live views of the plan, progress, git graph and spending

Every prompt, model, tool permission and budget can be edited

How it’s built

Claude Orchestra is built on TypeScript, Electron, git worktrees and Claude Code.

The repository is 94% TypeScript, 5% JavaScript by volume.

It carries 13 commits since Jul 2026, most recently changed in Jul 2026.

How the work is done

Every project here starts as a written spec, and the spec stays the source of truth once there’s code. The repository keeps a build plan with current status, an append-only decision log of where the code departed from the plan, and a handoff document that describes where things stand, including the unfinished parts.

That’s how one person keeps this many projects moving, and it’s what this page is written from.

Current state

Prototype. Works, but isn’t finished.

The source is private, so the way to check what this page says is to use the product.