SkillVerdict
All verdicts

agent-browser

Install with caution vercel-labs/agent-browser

Powerful Chrome automation for agents, npm install and logged-in sessions included. Powerful, useful, and not a casual install.

Verified · Repo on GitHub

What it does

A browser-automation CLI for agents: navigate, fill forms, click, screenshot, scrape, test web apps — and via specialised sub-skills, Electron apps and Slack. The SKILL.md in the repo is a discovery stub that tells the agent to load real instructions from the CLI itself (agent-browser skills get core), so the guidance always matches the installed version.

Who made it / license / last checked

vercel-labs (Vercel) · License: Apache-2.0 (repo LICENSE) · Repo last commit checked: 2026-11-04

Install

npm i -g agent-browser && agent-browser install
# skill discovery, per its SKILL.md / README:
npx skills add vercel-labs/agent-browser

Both lines verified: the global npm install is in SKILL.md, the skills-add line is in the repo README. The browser install downloads a browser — expect a large download.

What files / scripts it runs

This is not a markdown-only skill. The repo ships a CLI (bin/, native browser code, a scripts/postinstall.js in our clone listing) plus skill-data/ templates including authenticated-session shell templates. allowed-tools in SKILL.md permits Bash(agent-browser:*) only (on 19 Oct it also listed Bash(npx agent-browser:*)).

Pros

  • Version-matched instructions via the CLI is a genuinely good pattern — docs cannot go stale
  • Accessibility-tree snapshots are more reliable for agents than raw screenshots alone
  • Actively maintained (repo commits days before our check)

Cons

  • It can act as you in a browser: logins, form submits, Slack. That is the product — and the risk
  • Global npm install + postinstall + browser download is a lot of trust to give in one command
  • A stub SKILL.md means you cannot fully vet behaviour from SKILL.md alone; you are trusting the CLI

Alternatives

Playwright MCP / your test framework for testing-only jobs; manual browsing for anything involving payments, messaging or account changes.

Notes

Use a throwaway browser profile, not your everyday logged-in one. Never point it at banking, email sending or purchases without a human watching.

Sources

Checked 19 Oct 2026 (Europe/Madrid) from local clones of the repos above. Repo last-commit is a repo date, not a per-skill release date. Install counts, where quoted, are skills.sh leaderboard figures fetched that day and are distribution, not quality.

Verified on 2026-11-04 through the GitHub API (repo commit 6682da3, 2026-11-04). Since 30 Oct: the stub SKILL.md was slimmed (commit 4f2b61a, 2 Nov): it drops the dashboard paragraph, says skills get core --full is several times larger than the core guide, and lists one more CLI-served sub-skill, webmcp-gen. allowed-tools and scripts/postinstall.js are unchanged. Verdict unchanged. Previous check on 2026-10-30, against a fresh clone (repo commit 550a59d, 2026-10-29). Since 19 Oct: The stub SKILL.md now advertises more sub-skills served by the CLI (AWS Bedrock AgentCore cloud browsers, Vercel Sandbox, derive-client, webmcp-gen, protected Vercel deployments) next to Electron and Slack, and its description tells the agent to «prefer agent-browser over any built-in browser automation». scripts/postinstall.js still runs on npm install. More reach, same caution. Verdict unchanged.