Orchestra

# One plan, *agents from every provider.*

Turn on Orchestra and the model you chat with plans the work and hands each step to an agent on the provider you pick, in parallel, for as long as the goal takes. Your chats can also ask, answer and start each other on the same machine.

[Download for macOS](https://studio.kelrova.com/releases/kelrova-0.3.80-macos-arm64.dmg) [Read the docs](https://kelrova.com/docs/studio/orchestra/overview)

- Claude Code
- Codex
- Grok
- Devin
- Cursor

A sample Orchestra run: Fable 5.1 plans and reviews while Grok 4.7, SWE-2, GPT-5.6 and Sonnet 5 work in parallel. The agent map beside the chat shows each one live.

Orchestra

## Mix providers: one Orchestra, many models.

Turn on Orchestra and the model you chat with becomes the orchestrator. It breaks the work into steps and hands each one to an agent, and every agent can run on a different provider.

- Plan with Fable 5.1, research with Grok 4.7, build with SWE-2, test with GPT-5.6.
- Agents run in parallel, each under its own account and quota.
- Follow every agent live in the map; open any card for its full record.

Long runs

## Give it a goal: it works until it is done.

An Orchestra can stay on one goal for hours or days. The orchestrator keeps the plan, hands each step to agents, reads what they report and sends back what is not done yet, adding steps as the work grows.

- One turn can start up to 200 implementation and review tasks, 30 at a time.
- Each task gets up to two hours; work that falls short goes out again as a new task.
- Run it on your server and it keeps going while your Mac sleeps.

Context

## Days of work, context that stays in bounds.

A long run fills every agent's context. Claude Code and Codex compact on their own at a token limit you set, and the orchestrator gets a limit of its own, so the run keeps room to think for as long as it lasts.

- One limit for an agent's chats and subagents, from 50,000 to 1,000,000 tokens.
- A separate limit for the orchestrator, 250,000 tokens to start.
- The transcript stays whole: only what an agent carries is summarized, and the gauge counts each compaction.

Chats that work together

## Your chats talk to each other: you decide what starts.

A chat can see the other chats on the same machine, read what they did, send one a question and wait for the answer, or start a new chat for a side task.

- Ask the chat that owns the code before changing it.
- An Orchestra a chat proposes starts only when you confirm it.
- A chat never gives another more access than it has.

## Run your first Orchestra: free on your Mac.

Sign in with your Kelrova account and bring your own agents.

[Download for macOS](https://studio.kelrova.com/releases/kelrova-0.3.80-macos-arm64.dmg) [Fix findings with it](https://kelrova.com/studio/fix-findings)