Agent Browser
Token-efficient browser automation via accessibility snapshots (~200-400 tokens/page).
Artifacts
| Type | Detail |
|---|---|
tool | npm install -g agent-browser && agent-browser install |
skill | vercel-labs/agent-browser |
Requirements
- Chrome/Chromium installed
- Linux servers:
libasound2t64package +--no-sandboxflag (AppArmor)
What it does
Gives Claude browser control using accessibility tree snapshots instead of screenshots. Each page costs ~200-400 tokens (vs ~2000+ for screenshot-based approaches).
Capabilities:
- Navigate, click, type, scroll, wait
- Extract structured data from web pages
- Fill forms, interact with SPAs
- Multi-step workflows (login → navigate → extract)
Setup
- Ensure Chrome is installed
- Run
harness-kit add agent-browser- Installs
agent-browserCLI globally - Installs Playwright browsers via
agent-browser install - Adds the agent-browser skill via
npx skills add
- Installs
Notes
- Accessibility snapshots are text-based — no vision model needed
- Works with headless Chrome (CI/server environments)
- Skill source: skills.sh/vercel-labs/agent-browser