---
title: Check an app
description: Capture the app the agent built and judge what it shows.
---
> ## Documentation Index
> Fetch the complete documentation index at: https://docs.reeval.ai/llms.txt
> Use this file to discover all available pages before exploring further.

A file assertion can't tell you whether a page renders. The `nextApp` environment must contain the app project before you start. For more information about the `nextApp` environment, see [Environments](/configuration/environments).

Three calls check the app the agent built:

1. Start the app in the sandbox with `c.serve`.
2. Capture a page with `c.capture`.
3. Send the capture to the judge.

## Start a server

`c.serve` starts a command in the background and returns its port. reeval allocates a free port and sets it as `PORT`, so a server that reads `PORT` listens where reeval expects it. In the `evals/sign-in.eval.ts` file, build and start the app:

```ts title="evals/sign-in.eval.ts"
import { evaluate, expect } from '{{ pkg }}'
import { nextApp } from './environments.ts'

evaluate(
  'the sign-in page shows an email and a password field',
  { environment: nextApp },
  async (c) => {
    await c.send('Add a /login page with email and password fields.')
    await expect(c.exec('npm run build'), 'builds').toSucceed()
    const app = await c.serve('npm start')
    const shot = await c.capture(app, { path: '/login', wait: 'idle' })
    const q = 'Does the screenshot show an email field and a password field?'
    expect(await c.judge(q, { artifacts: [shot] }), 'two fields').toPass()
  },
)
```

## Check the running app

`c.exec` runs in the sandbox, so a request to `localhost` on the server's port reaches the app. `c.serve` returns as soon as the command starts, so retry until the server answers. In the `evals/sign-in.eval.ts` file, check the page over HTTP:

```ts title="evals/sign-in.eval.ts"
const url = `localhost:${app.port}/login`
await expect(
  c.exec(`curl -fsS --retry 10 --retry-connrefused ${url}`),
  'serves /login',
).toSucceed()
```

`-f` makes an HTTP error a non-zero exit code, which fails `toSucceed`. To assert on the body, read `stdout` from the `c.exec` result.

## Capture a page

`c.capture` takes a `ServeHandle` or a port and opens a path on it in a headless browser on the host through a tunnel to the sandbox. It waits until the server answers, then saves the image at the `artifacts/<name>.png` path in the attempt record. The sandbox image needs no browser, because reeval downloads one into its cache on the first capture.

`path` defaults to `/`. `name` defaults to the path, so `/login` saves `login.png`, and a name can't repeat in one attempt. `wait` defaults to `'load'`; use `'idle'` to wait until at most two requests are in flight for 500 ms, or `{ ms: 500 }` to wait a fixed time.

In the `evals/sign-in.eval.ts` file, name a second capture:

```ts title="evals/sign-in.eval.ts"
const error = await c.capture(app, {
  name: 'login-error',
  path: '/login?error=1',
})
```

To interact with the page before the capture, pass `instructions` as goal strings, which a vision model from `capture.model` drives.

## Judge a capture

The `artifacts` field on `c.judge` sends captures to the judge as images: the handles from `c.capture`, their names, or `true` for every capture so far. An evaluation model such as `typesafe-ai/jev-1.14.0` takes PNG captures, and a provider that can't take files fails the call. Ask what the picture shows, such as a field or a label, not whether it looks good.

In the `evals/sign-in.eval.ts` file, send both captures:

```ts title="evals/sign-in.eval.ts"
const q = 'Does the error page keep both fields and show one error message?'
expect(await c.judge(q, { artifacts: true }), 'error state').toPass()
```

For more information about the judge, its model, and its threshold, see [Judgments](/evaluations/grading/judgments).

In a comparison, a question shows each side's own capture when `artifacts` names the captures, or is `true`, on the question or the `question` option, so the model sees the baseline's page beside the variant's. For more information about pairwise questions, see [Ask questions across attempts](/evaluations/grading/questions).

## Set the viewport and browser

The viewport defaults to 1280×720. In the `evals/sign-in.eval.ts` file, set the capture width and height:

```ts title="evals/sign-in.eval.ts"
{ environment: nextApp, capture: { viewport: { width: 390, height: 844 } } },
```

`viewport: 'full-page'` captures the whole page. `browser: false` stops reeval from using its browser on the host, and `c.capture` then throws. For more information about every key, see [the `capture` option](/reference/options#capture).

## Reference

For more information about every argument `c.serve` takes, see [`c.serve`](/reference/attempt-context#cserve). For more information about every capture field, see [`c.capture`](/reference/attempt-context#ccapture). For runnable projects, see [`with-capture`](https://github.com/invertase/reeval/tree/main/examples/with-capture) and [`with-capture-journey`](https://github.com/invertase/reeval/tree/main/examples/with-capture-journey).