browser

Browser Automation

Automate browser interactions using the browse CLI with Claude.

Setup

Do not install packages during evals. The eval harness provides browse on PATH and pins it to the active benchmark session.

Environment Selection

The eval harness pins browse to one isolated session and preselects the intended environment.

• Local mode uses local Chrome.
• Remote mode uses Browserbase when credentials are configured.
• Do not switch environments unless the task or harness explicitly requires it.

Page State

Prefer browse snapshot over screenshots. It returns a structured accessibility tree with element refs that can be used for interactions.

Only use screenshots when visual layout or image content is necessary.

Useful state checks:

browse status
browse get url
browse get title
browse snapshot

Interaction Workflow

Typical flow:

1. Navigate to the starting URL with browse open <url>.
2. Run browse snapshot to inspect structure and element refs.
3. Interact with refs from the snapshot, selectors, or keyboard actions.
4. Run browse snapshot or another state check to verify the result.
5. Repeat until the benchmark instruction is complete.

Common interaction patterns:

browse click <ref>
browse type <text>
browse press <key>
browse fill <selector> <value>
browse wait timeout <ms>

Best Practices

• Always navigate before interacting.
• Use refs from browse snapshot when possible.
• Verify each important action before moving on.
• Keep actions focused on the benchmark instruction.
• Do not edit repository files.
• Do not use browser/network tools outside the provided browse command.

Completion

When finished, report the benchmark result in the exact EVAL_RESULT format requested by the harness prompt.