Tulum DAO

Run a fleet of coding agents as a team.

OrchestraOS is an open harness, published by Tulum DAO. Agents message each other, remember across restarts, rotate themselves before they run out of context, and put every real decision in front of you on your phone.

github.com/Tulum-DAO/orchestraos

Already have a server and a terminal open? New to this? Start from scratch.

  1. Log in to one agent CLI first

    Install Claude Code, Gemini (the agy CLI) or Codex, and finish its login before you clone. A seat started against a CLI you haven't logged in to fails without telling you why. Prerequisites and login commands

  2. Clone and install

    git clone https://github.com/Tulum-DAO/orchestraos.git orchestraos && cd orchestraosmake install

    Then orchestra init --yes (about five minutes) and orchestra doctor. Read the full install

Build with us

OrchestraOS is built in the open, and the community is always making contributions to the build. It runs one operator's fleet every day, and that setup is the reference install. Pick a track or a good first issue and add yours.

Seats and generations

Each agent is a seat. Before its context runs out, it hands off to a successor that proves it read the handoff. How rotation works

An agent is a seat with a lineage. Each generation is one CLI session: Claude Code, Gemini or Codex. When a generation nears its context ceiling it writes a handoff, a successor boots, proves it understood by answering questions anchored in its predecessor's own state, and is promoted.

Rotation is on by default for Claude seats. Gemini and Codex are experimental and stay unarmed until you opt in. How rotation works

Smart inter-agent messaging that doesn't interrupt work

Agents hand each other work. Mail for a busy agent waits until its turn ends.

Agents hand each other work through a durable message store with inboxes, acknowledgements and threads. Mail for a busy seat is parked and delivered when its turn ends, so no agent is cut off mid-task and nothing is lost.

  • task
  • reply
  • task_complete
  • escalate

Memory that survives a restart

Facts and each seat's memory outlive every restart and rotation.

A shared facts store with freshness, memory directories for each seat that outlive rotation, and a recall hook that feeds Arturo, the assistant.

Every real decision comes up to you

Anything an agent needs from you is a card on your phone, watch or dashboard. Your answer resumes it.

Anything an agent needs from you is a card: approve or deny, a menu, a questionnaire, or a "waiting on you" block. Cards show on the web dashboard and on the iOS and watch app, and your answer resumes the agent that asked.

The runtime catalog checks which CLIs you have installed and logged in to, so the harness runs on a subscription you already pay for.

An assistant that sees what you see

Ask Arturo from any dashboard page. It focuses on the page you are on and reads its live data when you ask.

The Ask Arturo pill sits bottom-right on every dashboard page. Its chat opens beside your work without dimming or blocking it, and it focuses on the page you are looking at: a card in the composer says which page, and its × turns that off. Ask “what am I looking at?” and Arturo reads that screen's live data.

Talk or type. Conversations are kept, so you can switch back to one and pick up where you left off. Arturo sits on top of your agents: it can see the team, answer cards and message seats, within the scopes you give it, and when you ask.

Two ways to run it

How to contribute

Fork the repo, branch, and open a pull request against main. CI runs the tests for each package and a secrets scan, and both must pass. Sign off your commits (DCO). There are thirteen tracks, each with its own doc and a start prompt you can paste into your own agent, and a list of small good-first issues.