# Runbook for AI agents

> How an agent connects itself to flurry and gets started.

Canonical: https://dvt.dev/docs/agents/

You are an AI agent and a person asked you to connect to flurry. Follow these steps in order.

1. **Ask first.** Tell the person you're about to add flurry to your tools and wait for a yes. Never ask for a token in chat, and never paste one there.
2. **Find the address.** If you were given a flurry URL, read `<url>/connect.md`: it lists the exact command for each host at that address. Without a URL, use `http://127.0.0.1:8787` for flurry on this machine.
3. **Connect.** If you can run shell commands and you are Claude Code or Cortex Code, run your host's command from [connect](https://dvt.dev/docs/connect/) yourself. Otherwise, give the person the clicks for their app from the same page.
4. **Start.** Call `flurry_reference()` with no arguments. Its `start` block says who you are, where you can build, three example apps to open, and prompts to try.
5. **Report.** Tell the person: "Connected as <user>. Here are 3 things to try," using `start.first_prompts`. Then read `flurry_reference(topic='workflow')` before you build anything.

## Rules that save a round trip

- Every change takes a specific change note of at least three words. "update" and "wip" are refused.
- Mutating tools support `preview=true`. Preview first and show the person what will change.
- Drafts never go live on their own. Tell the person to review and publish on the app's **Versions** page.
- If a call fails, read the error code and see [troubleshooting](https://dvt.dev/docs/troubleshooting/).