The Orchestrator

What It Is

When you open Orcha, the home screen is a chat. Behind it is one persistent Claude session with tools over your whole workspace — it can list your sessions, read their transcripts, dispatch new agents, and steer running ones. It runs on your Claude subscription, not an API key.

The conversation persists across app launches. Close Orcha, come back tomorrow, and the orchestrator picks up where it left off.

Ask What's Running

Ask “what's running?” or “where did the api session leave off?” and the orchestrator checks the actual state of your sessions — not a cached guess. It distinguishes an agent that's working from one that's waiting on you or blocked on a permission prompt, and it quotes what each one is actually doing as evidence.

It can read the tail of any session's terminal and transcript, so “summarize what the billing agent did” works even for sessions you never watched.

Dispatch & Steer

Tell it to do something — “fix the webhook bug in the api project” — and it dispatches a background agent into that project. When the agent finishes or gets stuck, the status posts back into the conversation. You never have to go hunting for a terminal.

For a session that's already running, it can steer: send a course correction into the live session mid-flight — this works for chat sessionsexactly as it does for terminal sessions; chats join the fleet. Sessions the orchestrator owns are kept out of your Needs You inbox — it reports on them in the conversation instead.

Build by Asking

• Automations— “Every morning at 9, run the tests and summarize what broke.” The orchestrator interviews you for the details, then the automation exists. See The Library for managing them.
• Projects— “Set up a new project called billing-service.” It creates the folder, registers the project, and it appears in your workspace exactly as if you'd used New Project.
• Backlog— click any backlog item and it lands in the orchestrator's composer, ready to dispatch.

Models & History

The composer has a model picker, just like Claude chat — switch models whenever you like and the conversation continues. The history button lists your past conversations; start a new one anytime and switch back later without losing anything.

What It Can't Do

The orchestrator has no shell or file access of its own — it works entirely through its session tools. It can't edit your files directly; it dispatches agents that do. That boundary is deliberate: the orchestrator coordinates; your sessions execute— chat or terminal. It can see and steer both, but the work itself always happens in a session, never in the orchestrator's own hands.