# Quickstart


Use this when you want the fastest hands-on ZCP loop: deploy an **AI Agent** recipe, authorize Claude Code, open Browser VS Code, and ask for one product change.

Some [Zerops recipes](https://app.zerops.io/recipes) include an **AI Agent** environment. That preset creates app services, managed services, the `zcp@1` workspace, Browser VS Code, and bundled agent wiring in one deploy.

This quickstart uses [Laravel showcase](https://app.zerops.io/recipes/laravel-showcase?environment=ai-agent) because it includes real managed services. The same flow works for other recipes that offer AI Agent (local variants exist too).

## Prerequisites

- A Zerops account with permission to create a project.
- A Claude Code subscription login or API authentication. Zerops wires Claude Code to ZCP, but your agent subscription or model credentials stay yours.

## 1. Choose the AI Agent recipe

1. Open the [Zerops recipes catalog](https://app.zerops.io/recipes).
2. Open a recipe with an **AI Agent** environment. For example, open [Laravel showcase](https://app.zerops.io/recipes/laravel-showcase?environment=ai-agent).
3. Select **AI Agent**.
4. Keep **Coding Agent** and **Cloud IDE** enabled.
5. Deploy the recipe.

The deploy creates app runtimes, managed dependencies, and a `zcp@1` workspace. In this recipe it appears as the `zcp` service. The agent, terminal, and browser IDE run there. App code still deploys to the app runtimes, not to `zcp`.

## 2. Authorize Claude Code

When provisioning finishes, the dashboard opens the Claude Code authentication flow. Use your own Claude Code subscription login or API credentials.

This is separate from `ZCP_API_KEY`, the Zerops token used by ZCP. Your Claude login or API key is used only by the bundled agent.

If you close the prompt, open the `zcp` service in the dashboard. Its panel shows the browser workspace, web terminal, SSH access, desktop editor access, and agent authorization state.

## 3. Open the workspace

After authentication, continue into **Browser VS Code**. The workspace opens with files, ZCP configuration, terminal access, and the Claude Code panel.

You are now inside the remote workspace. The agent can read live state, use the MCP tools, reach private services, and deploy app changes to the runtimes created by the recipe.

## 4. Ask for a product outcome

In Claude Code, ask for the app behavior in natural language. A good first prompt is intentionally short:

```text
Build a task board for my team.
Tasks should stay saved after refresh.
```

The agent should deploy, verify, read evidence, and return proof for the app task. Add detail only when it changes the work.

Add details when they change the product, stack, runtime layout, acceptance criteria, or delivery preference:

```text
In this Laravel app, add a task board backed by the existing PostgreSQL service.
A user can create a task, refresh the page, and still see it.
```

:::note Prompt shape

Quickstart prompts are short so you can see what ZCP carries behind the request. Longer, detailed prompts are fine for larger engineering work, especially when they change product behavior, stack, runtime layout, acceptance criteria, delivery preference, credentials, or safety approvals.
:::

## 5. Read the proof

The final answer should give you a real URL, endpoint, UI state, stored result, or blocker. Open the URL and try the behavior the agent says it verified.

For the task-board prompt, create a task, refresh the page, and confirm the task is still there. If it disappears, the app is not done; ask the agent to verify that exact behavior again and continue from current evidence rather than starting over.

If the agent cannot finish, useful output names the blocker: missing credential, missing decision, unsupported runtime choice, or repeated failure it could not recover from.

## Next steps

- [Build and ship](/zcp/workflows/build-with-zcp) — Runtime layout, development work, delivery, packaging, and production release.
- [How it works](/zcp/concept/how-it-works) — Live state, runtime fit, app wiring, deploy evidence, behavior proof, and blockers.
- [Trust model](/zcp/security/trust-model) — What the project-scoped token lets the agent do, and where the safety boundary sits.