Agent Browser
Headless Chromium browser for web browsing and screenshots inside the sandbox.
A headless Chromium browser for web browsing and screenshots. Uses the agent-browser binary from Vercel Labs.
| Property | Value |
|---|---|
| ID | agent-browser |
| Version | 0.16.3 |
| Binary | ~/.local/bin/agent-browser |
Installation
After the binary is downloaded from GitHub releases, it runs agent-browser install --with-deps to fetch Chromium and required system libraries.
Required network domains:
github.comrelease-assets.githubusercontent.comcdn.playwright.devstorage.googleapis.comcdn.amazonlinux.com
If your sandbox uses a restricted network policy, the deploy CLI will prompt you to add these domains to the allowlist.
Environment variables
ClawRun automatically injects these environment variables into the agent daemon when Agent Browser is enabled:
| Variable | Value | Purpose |
|---|---|---|
AGENT_BROWSER_NATIVE | 1 | Uses the built-in Rust daemon instead of looking for daemon.js from the npm package. Required for the standalone binary in sandboxes. |
AGENT_BROWSER_MAX_OUTPUT | 50000 | Caps snapshot output to 50,000 characters. Prevents context overflow on dense pages. The default (unlimited) can produce 64K+ characters. |
These are set via the tool's runtimeEnv and do not need manual configuration.
Skill file
Agent Browser ships a SKILL.md that is seeded into the agent's workspace at workspace/skills/agent-browser/SKILL.md during deploy. It contains detailed instructions for browser automation, including navigation, screenshots, form interaction, and output handling.