Your first AI worker

This page walks you from nothing to an AI agent that picks up tickets in your project and works them on its own - claiming, tracking time, committing, and closing like any other team member. No prior experience with terminals or agents needed; every step tells you exactly where to click and what to paste. Ten minutes, start to finish.

You can run one worker (it implements tickets), or a duo: one worker plus one reviewer that checks the finished work - reading the changes, running the tests, and visually inspecting screenshots before approving.

What you need

• A computer running macOS or Linux (on Windows, use WSL). The agent works on this machine, so it should have the project repository on it - or be allowed to clone it.

• An orboto admin login (you'll create a bot account for the agent).

• An AI brain for the agent - one of:

  • an Anthropic API key (recommended: sign in, open API Keys, click Create key, copy the sk-ant-... key), or
  • an Ollama server (local, or Ollama Cloud with an API key), or
  • LM Studio running a local model.

  For Ollama and LM Studio the wizard installs the needed pi extension (pi-lm-providers) for you.

• Node.js 20 or newer. Not sure? You'll find out in step 3 - and if it's missing, download the installer from nodejs.org and click through it like any other app.

Step 1: create the agent's account in orboto

Agents get their own account so their work stays clearly attributable - and so you can pause them anytime.

1. In orboto, open Admin → Multi-Agent. At the top you'll find the Guided setup: parallel agents card.
2. Create a service account - give it a name and email people will recognize, e.g. "Worker Bot" / worker-bot@yourteam.example.
3. Assign the agent role - pick the built-in service-account role the wizard recommends.
4. Mint the API key - name it (e.g. "worker key") and copy it immediately: it starts with orb_ and is shown only this once.

Planning the duo? Run the same wizard a second time for the reviewer (e.g. "Review Bot" / review-bot@yourteam.example) and copy its key too.

Step 2: open a terminal

• macOS: press Cmd + Space, type Terminal, press Enter.
• Linux: open your Terminal app.
• Windows: open your WSL/Ubuntu terminal.

Then move into the project folder the agent should work on - type cd, a space, then drag the project folder from your file manager into the terminal window, and press Enter.

Step 3: run the setup script

Paste this line and press Enter:

bash <(curl -fsSL https://docs.orboto.io/runner-setup.sh)

Nothing is left behind on your disk - the installer runs straight from the line above (only your answers are saved, see below).

The wizard walks you through seven short steps - what to run (worker or duo), your orboto address, the bot key(s), the project, and the AI provider (Anthropic, Ollama, or LM Studio - extensions are installed for you). Every value is checked immediately, every step explains where to find what it asks for, and at any question you can type b to go back a step or q to quit. Nothing starts before a summary you confirm.

The script installs what's missing (the orboto command-line tool and the pi coding agent) and then asks you a few questions - paste your answers and press Enter after each:

Answers are stored (safely, readable only by you) in ~/.orboto-runners/config.env - the next start skips the questions.

Step 4: watch it work

That's it - the agent is running. It loads your workspace's agent rules, asks orboto for its next ticket, and starts working. Whenever it's idle for five minutes, it pulls the next one. In the duo, the reviewer picks up tickets your worker finishes and reviews them - with a vision-capable model, so it actually looks at screenshots and UI output.

• See what it's doing: the terminal shows the live log - and inside orboto, the agent appears in Admin → Multi-Agent (live presence) and on every ticket it touches, like any teammate.
• Give it directions while it runs, from any other terminal: orboto agent-notify worker-bot@yourteam.example "Do ACME-42 first" --project ACME
• Pause it in Admin → Multi-Agent, or stop it with Ctrl+C in the terminal. Starting the script again resumes the same session with its memory intact.
• In the duo with tmux installed, both agents run in one session: attach with tmux attach -t orboto-runners, switch windows with Ctrl+B then 0/1, detach with Ctrl+B then D.

Choosing different models

The worker and reviewer default to Anthropic's Sonnet (vision-capable). To use something else, set the model before starting:

IMPL_MODEL='openai/gpt-5.2' REVIEW_MODEL='anthropic/*sonnet*' bash <(curl -fsSL https://docs.orboto.io/runner-setup.sh) duo

Keep the reviewer on a vision-capable model - that's what lets it judge screenshots, not just code.

When you outgrow this

The same runner scales to always-on, specialized lanes on a server - see Run an agent fleet. A runner set up here is not locked to the project or model you started it with: register it as a lane and you can re-target it - other projects, another role, a different model, paused or running - from Admin → Agent Lanes in orboto, and the change reaches the running agent without touching the machine it runs on Added in v0.176.0. And if you'd rather wire it up by hand (no script), the fleet page's smoke-test section shows the underlying orboto pi-runner command.