browsing-skills

Browsing Skills: turn website actions into structured APIs

Make your AI agent call web application actions like APIs, even when the website does not expose the API you need.

browsing-skills is an open-source library of website-specific browser actions for AI agents. Each supported website is packaged as a standard SKILL.md skill with focused actions, documented inputs, and structured outputs.

Why

Have you ever tried to point an AI agent at a web app you use every day, only to watch it crawl through the page, inspect huge DOM dumps, guess selectors, miss fields, and burn tokens?

Browsing skills make common web actions fast and repeatable:

• No selector research loop. The agent loads the right action reference instead of rediscovering the page from scratch.
• One page execution call. After navigation, the agent runs maintained action code with page.evaluate() or Chrome Bridge /run-action.
• Lower token and time cost. Less DOM reading, fewer repair loops, fewer failed extraction attempts.
• Cleaner results. Actions return structured data in a predictable shape.

We benchmark each action against a no-skill browser agent and track both time and tokens in BENCHMARKS.md. In one Booking.com search benchmark, the skill path used 3,903 tokens and finished in ~9.5s; the no-skill DOM-inspection loop used 49,290 tokens and took ~82.5s.

Supported Websites

The library currently includes skills for:

Browsing skills are most useful for web applications that do not expose the API an agent needs. Sometimes that is intentional, as with many social platforms; sometimes the product simply has an old, incomplete, or unavailable API. In those cases, the browser UI is the integration surface, and a skill turns that UI into a repeatable action.

Two ways to install

A. The umbrella (I want all supported sites)

Point your agent at the top-level SKILL.md. It tells the agent which sites are supported and how to load each site's skill on demand:

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

This is the right choice if your agent might interact with any of the supported sites.

B. A single site (I only care about one)

If you're building an agent that only ever works with one site, skip the umbrella — install that site's SKILL.md directly. For example, a LinkedIn-only agent:

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

This is leaner — no list of unrelated sites, no routing step. Each site is a complete, standalone skill.

Repo layout

browsing-skills/
├── SKILL.md                     # umbrella — lists supported sites, routes agents
├── skills/
│   ├── linkedin.com/
│   │   ├── SKILL.md             # linkedin action index
│   │   └── references/
│   │       └── post-data.md     # one action spec + JS
│   └── x.com/
│       ├── SKILL.md             # x (twitter) action index
│       └── references/
│           ├── post-data.md     # one action spec + JS
│           ├── profile-data.md  # one action spec + JS
│           └── search.md        # one action spec + JS
├── chrome-bridge/               # optional browser-access companion
└── tools/                       # validation + umbrella regeneration

A site's SKILL.md is a compact action index. Each action's requirements, navigation instructions, code block, and return shape live in a separate file under skills/<domain>/references/, so agents can load only the action they need. Adding a new action means updating the index and adding one action-specific reference file.

How actions run

Most modern sites are JavaScript-rendered and require a real browser. Use your agent's browser integration, Playwright, or the optional chrome-bridge/ companion to run actions inside Chrome.

Reference files contain self-contained JavaScript action objects. Once the target page is loaded, run the action in the page context with page.evaluate() or Chrome Bridge's /run-action endpoint:

var result = await page.evaluate(async function(code) {
  var tool = eval(code);
  return await tool.execute({ mode: "data" });
}, scriptCode);

Actions return:

{ content: [{ type: "text", text: "..." }] }

The text value is usually JSON, but some actions also support mode: "display" for self-contained HTML output.

The action object format is Web MCP aligned: each action declares a name, description, input schema, and executable handler. That means the code in these reference files is not only useful as an external skill. A website can also adopt the same action definitions directly in its own frontend or agent endpoint, exposing first-party agentic browsing support without requiring agents to install a separate skill for that site.

Contributing

To add a new supported site or improve an existing one:

1. Fork the repo, create a branch.
2. Add or edit skills/<domain>/SKILL.md and skills/<domain>/references/*.md.
3. Test your code against a real page (via chrome-bridge or any browser automation you have).
4. Open a PR. CI validates frontmatter, JS syntax, and structure.
5. On merge to main, the umbrella SKILL.md's supported-site list regenerates automatically.

Site SKILL.md template

---
name: browsing-<site>
description: "Use when the user wants to interact with <site> — <list the actions>. <auth/browser notes>."
---

# <site> — Browsing Skill

Use this index to choose the <site> action that matches the user request, then open only the linked reference file for the complete navigation, requirements, code, and return shape.

## Action Index

- **<action-1>** — Short explanation of when to use this action. Full spec: [references/<action-1>.md](references/<action-1>.md).
- **<action-2>** — Short explanation of when to use this action. Full spec: [references/<action-2>.md](references/<action-2>.md).

Put each full action specification in its own file under skills/<domain>/references/:

# <site> — <action-1> Reference

## Requirements

Auth notes, browser notes, cookie injection snippets, etc.

## How to run this action

Shared execution pattern (page.evaluate / chrome-bridge `/run-action`).

---

## Action: <action-1>

**Navigate to:** `https://...`

**Code:**

```js
({
  name: "<site>-<action-1>",
  description: "...",
  inputSchema: { ... },
  execute: function(params) { /* ... */ }
})
```

**Returns:** `{ ... }`

Conventions

• One SKILL.md index per site. The site SKILL.md lists available actions with short explanations and links to full specs under references/.
• One reference file per action. Put navigation instructions, auth/browser requirements, the executable js block, and the return shape for one action in skills/<domain>/references/<action>.md.
• The description in frontmatter is the trigger. Agent frameworks match user intent against it. Be specific about what the skill does and when.
• Action code blocks (({ name, execute, ... })) are validated by CI for JS syntax. Non-action snippets (like cookie-injection examples) aren't validated — they're documentation.
• Use var (not let/const) for maximum compatibility in any browser.
• Self-contained code — no external imports, no CDN scripts.
• Action object format — each action is an object with name, description, inputSchema, and an execute(params) function. It returns { content: [{ type: "text", text: ... }] }.

Security and privacy

• Do not commit cookies, API keys, session tokens, or personal browser profile paths.
• Keep examples generic. If an action requires authentication, document the requirement without including real account data.
• Chrome Bridge runs code in real browser tabs. Review action code before running it on logged-in sessions, and keep write actions explicitly opt-in.

Responsible Use

This project provides code and documentation for browser automation. You are responsible for how you use it.

• Follow the terms of service, robots policies, rate limits, and other rules of any website you access.
• Only automate accounts, pages, and data you are authorized to use.
• Do not use these skills to bypass access controls, paywalls, security controls, or privacy settings.
• Be careful with logged-in sessions. Automated browsing can trigger fraud, abuse, or bot-detection systems, and websites may restrict, suspend, or block accounts.
• The maintainers are not responsible for account restrictions, data access issues, legal claims, or other consequences from your use of these skills or Chrome Bridge.

Reporting broken or missing skills

Open a GitHub issue using one of the templates:

• skill-broken — a skill stopped working
• skill-request — a site not yet covered
• skill-enhancement — existing skill needs more actions or fields

License

MIT — see LICENSE.