Skip to content

Built-in Tool: Browser

Document ID: TRT-108
File Path: docs/07-tool-runtime/browser.md
Version: 1.0.0
Status: Draft
Owner: AI Platform Team
Last Updated: 2026-06-27

1. Purpose

The browser.* built-in tools give agents a headless browser for tasks that need rendered pages — scraping JS-heavy sites, filling forms, capturing screenshots. High-resource and high-risk; sandboxed strongly.

2. Operations

Tool	Description
`browser.navigate`	Open a URL, return rendered text/DOM
`browser.click`	Click an element
`browser.fill`	Fill form fields
`browser.screenshot`	Capture a screenshot
`browser.extract`	Extract structured content via selectors

3. Schema (example: `browser.navigate`)

// input
{ "url": "https://example.com/docs", "wait_for": "networkidle", "timeout_ms": 30000 }
// output
{ "url": "...", "title": "...", "text": "...", "status": 200 }

Screenshots return an artifact reference stored in object storage, not inline bytes.

4. Permissions

net:egress:<domain>     (per allowed site)
browser:use

Egress is restricted to granted domains; broad browsing requires explicit, broad grants flagged as high-risk.

5. Sandbox & Safety

The browser runs in a strong sandbox (gVisor/microVM) on the untrusted worker pool due to its attack surface (backends).
Egress allowlist + DNS restriction prevent navigation to disallowed hosts.
CPU/memory/time limits are higher than other tools but still enforced; sessions are ephemeral.
No credentials are auto-filled unless provided as secret references.

6. Determinism & Caching

Browsing is non-deterministic (live web) and not cached. For reproducible tests, point at fixtures or recorded pages.

7. Example

wovyr tools enable browser.navigate --project research
wovyr tools invoke browser.navigate --input '{"url":"https://example.com"}'

9. Revision History

Version	Date	Description
1.0.0	2026-06-27	Initial Browser tool spec