Browsing Skills

browsing-skills

An open-source library of website-specific browsing skills for AI agents. Each supported website is a first-class skill in this repo — the umbrella (this file) just lists what's available and tells you how to load the one you need.

Repo: https://github.com/browsing-skills/browsing-skills

When to use this umbrella

You're loaded into an agent that might work with any website. When the user mentions a specific site, you want to check: is there a prebuilt skill for it?

• Supported site (<!-- DOMAINS:START -->airbnb.com, aliexpress.com, amazon.com, booking.com, capterra.com, craigslist.org, facebook.com, g2.com, glassdoor.com, instagram.com, linkedin.com, madlan.co.il, maps.google.com, news.ycombinator.com, reddit.com, substack.com, target.com, tiktok.com, walmart.com, x.com, yad2.co.il, zillow.com<!-- DOMAINS:END -->) → load the site's skill, follow its instructions.
• Unsupported site → handle the request the way you normally would.

When to skip this umbrella

If your agent only ever interacts with one site (e.g. you're building a LinkedIn scraping tool), install that site's skill directly and forget this umbrella exists:

Copy code
https://raw.githubusercontent.com/browsing-skills/browsing-skills/main/skills/linkedin.com/SKILL.md

How to load a site's skill

Fetch the raw action index for the domain the user mentioned:

Copy code
GET https://raw.githubusercontent.com/browsing-skills/browsing-skills/main/skills/<domain>/SKILL.md

Examples:

• skills/x.com/SKILL.md
• skills/linkedin.com/SKILL.md

Read the fetched skill index, choose the matching action, then fetch only that action's linked markdown file under skills/<domain>/references/. The reference file tells you:

• Which URL to navigate to for the action
• The JavaScript to execute, in a js code block
• The action's return shape
• Auth and browser requirements

How site skills work (mental model)

A site skill is a small markdown index with:

1. Frontmatter — machine-readable metadata (name, description, etc.)
2. Action Index — short descriptions of available actions and links to full specs
3. Action reference markdown — one file per action under references/, with requirements, navigation details, a fenced js code block, and the return shape

Benchmarks

This repo benchmarks each action against a no-skill browser agent. The skill path loads the site action index plus one action reference, then runs the maintained action code. The no-skill path inspects the live page DOM, derives selectors at runtime, and iterates until it succeeds or times out.

Each site SKILL.md includes a compact benchmark table for its actions. Detailed benchmark notes and methodology live in BENCHMARKS.md.

You don't need any custom tooling — fetch the file, read it, follow it.

Browser access

Most skills require a real browser because target sites are JS-rendered. If your agent has browser access (Playwright, a Chrome extension, etc.), use it.

If you don't have browser access, the optional Chrome Bridge companion is a tiny Chrome extension + local bridge that lets you run skills in the user's real Chrome tabs. Setup: under a minute.

Chrome Bridge Quickstart

Use Chrome Bridge when the target site requires the user's already-signed-in Chrome session, or when Playwright/new browser contexts cannot access the page.

Start the local bridge server from this repo:

bashCopy code
cd chrome-bridge/server
node bridge.js

Expected server output:

textCopy code
[bridge] Listening on http://127.0.0.1:7865
[bridge] Waiting for extension to connect...
[bridge] Extension connected

If the extension is not connected, load chrome-bridge/extension/ as an unpacked Chrome extension from chrome://extensions/.

Check open tabs:

bashCopy code
curl http://127.0.0.1:7865/tabs

Run a skill action in a specific tab:

bashCopy code
curl -X POST http://127.0.0.1:7865/run-action \
  -H 'Content-Type: application/json' \
  -d '{"tabId":123,"code":"({ execute: function(params) { return { content: [{ type: \"text\", text: document.title }] }; } })","params":{"mode":"data"}}'

When using Bridge on sensitive logged-in accounts, keep actions read-only unless the user explicitly asks otherwise. Prefer selecting a tabId from /tabs instead of relying on the active tab.

Reporting issues / requesting skills

• A skill is broken: https://github.com/browsing-skills/browsing-skills/issues/new?template=skill-broken.md
• A site isn't supported yet: https://github.com/browsing-skills/browsing-skills/issues/new?template=skill-request.md
• An existing skill needs more fields: https://github.com/browsing-skills/browsing-skills/issues/new?template=skill-enhancement.md

Contributing

See the README.