anysite-mcp

Your agent learns a website once, then keeps it as a tool.

A real Chrome, signed in as you, on your own computer.

claude mcp add anysite -- npx -y anysite-mcp

No API keys. No cloud browser. No extra LLM. Your logins never leave your machine.

Quick start  ·  Why it's different  ·  Tools  ·  Shortcuts  ·  Shortcut Library  ·  Packs  ·  Safety  ·  FAQ

What it is

anysite-mcp is an MCP server that drives a real Chrome on your computer. Every site gets its own browser profile that you sign in to once, by hand. The agent then works the site the way a screen reader does, by role and name, so small redesigns don't break it.

When the agent gets a job done, it can save the steps as a shortcut, and that shortcut becomes a new tool it can call in one step next time.

See it in action

You:    Ask Gemini to summarise this PDF, then add the summary as a source to my "Thesis" notebook in NotebookLM.

Claude: gemini_ask { prompt: "Summarise this PDF in 10 bullet points", files: ["~/Downloads/paper.pdf"] }
        → - The study compares …   (Chat: https://gemini.google.com/app/…)
        notebooklm_add_source { notebook: "Thesis", text: "- The study compares …" }
        → Added pasted text.

You:    Check my GitHub notifications.

Claude: web_open { site: "github.com", url: "/notifications" }
        → This looks like a sign-in page. Call web_login for "github.com" …
        web_login { site: "github.com" }
        "A Chrome window opened on GitHub. Sign in there, then quit that window and tell me."
You:    done
Claude: web_open { site: "github.com", url: "/notifications" }   → (accessibility tree of your inbox)
        web_save_shortcut { site: "github.com", name: "notifications",
          description: "List unread GitHub notifications",
          steps: [{ "action": "open", "url": "/notifications" }, { "action": "read", "role": "main" }] }
        → Saved github_com_notifications.

(Next week, in a new session)
You:    Anything new on GitHub?
Claude: github_com_notifications {}   → 3 unread: …

Why another browser MCP?

Most browser tools make the agent rediscover a site every session. anysite-mcp keeps what it learned.

[!NOTE] Profiles are kept apart on purpose. Copying a signed-in profile so two browsers share one session gets that session revoked by Google (and other careful sites), which logs you out everywhere.

Quick start

Requirements: Node 22+ and Chrome (or npx playwright install chromium). Works on macOS, Windows and Linux.

1. Sign in to the sites you want, once each. A normal Chrome window opens; sign in, then quit it (Cmd+Q on a Mac, close all its windows on Windows and Linux).

npx -y anysite-mcp login gemini
npx -y anysite-mcp login notebooklm
npx -y anysite-mcp login github.com

2. Add the server to your agent.

claude mcp add anysite -- npx -y anysite-mcp

{ "mcpServers": { "anysite": { "command": "npx", "args": ["-y", "anysite-mcp"] } } }

3. Ask your agent to do something on one of those sites. That's it.

With no arguments the server offers every site you have signed in to. To pick sites, list them: npx -y anysite-mcp gemini github.com. The agent can also add a new site at any time by calling web_login.

anysite-mcp is listed in the MCP Registry as io.github.DwytMCL/anysite-mcp, so clients that browse the registry can find and install it.

Tools

Any site

Each tool takes site: a pack name like gemini, or a domain like github.com.

Every tool carries MCP annotations for clients that use them, for example to ask before a destructive tool runs. Tools that only read are marked read-only, and tools that can change something on a page (clicking, typing, choosing, pressing keys, uploading) are marked destructive. Shortcut tools are marked by what their steps do.

Packs included

Shortcuts

A shortcut is a list of steps, each one the arguments of a web_* tool plus its action. {{param}} is filled in from the tool's arguments, and the shortcut returns the text of its read steps.

• A step's expect is text that must be on the page afterwards, so a shortcut fails instead of returning the wrong thing.
• Before saving, the steps are run once with test_args (set verify: false for steps that must not happen twice, like sending).
• Shortcuts are saved in ~/.anysite-mcp/sites/<site>.json, which you can edit, share or commit. The file is versioned ("version": 1); an older file is kept as <site>.v0.bak the first time it is saved.

{
  "search": {
    "description": "Search Hacker News and return the results",
    "params": { "query": "What to search for" },
    "steps": [
      { "action": "open", "url": "https://hn.algolia.com/?q={{query}}" },
      { "action": "wait", "seconds": 5, "until_text": "results", "expect": "{{query}}" },
      { "action": "read", "role": "main" }
    ]
  }
}

Params can also be { "description", "type": "number" | "boolean", "optional", "default" }.

When a site changes

A step whose element is gone is repaired before it fails:

1. First with other ways to find it that were noted when the shortcut was saved (its label, its text).
2. Then with the closest element of a similar role on the page. "Search products" finds "Product search"; a button that became a link is still found.

Only a single clear match counts. The shortcut file is updated and the agent is told what changed. If nothing fits, the error says which step and why, and the agent can redo the job and save the shortcut again.

Record a shortcut by doing the job yourself

npx -y anysite-mcp record github.com star_repo

Chrome opens on the site; click and type as usual, then press Enter in the terminal. Each typed value becomes a parameter, with what you typed as its default. The agent can do the same with web_record.

Check your shortcuts

npx -y anysite-mcp doctor

doctor follows each shortcut while its steps only open and read pages, checks that the element of the first step that would act is there, and repairs what it can. It never clicks or types.

Each parameter gets its default if it has one. Otherwise an optional parameter is left empty, so the check sees the page the way a plain call would, and a required one gets "test". If an optional parameter breaks the page when it's empty (say, an empty search in the page address), give it a default. Note that this also changes what users get when they leave it out.

The Shortcut Library

Shared shortcuts live in library/, and you can browse them on the Shortcut Library site. Add a site's shortcuts with one command, and send yours in a pull request.

npx -y anysite-mcp add github.com                    # install a site's shared shortcuts
npx -y anysite-mcp library search wikipedia          # matches, with their health and add command
npx -y anysite-mcp library list github.com           # every site, or one site's shortcuts
npx -y anysite-mcp library info en.wikipedia.org/summary
npx -y anysite-mcp library check                    # which of your shortcuts differ from the library's

Health checked every week. A scheduled job runs every library shortcut against the live site, and each one shows its result on the site and in the terminal:

library, add and permissions don't need Chrome installed.

Write a pack

For logic that steps can't express (waiting for a streamed answer, parsing a list), write a pack: a file in ~/.anysite-mcp/sites/<name>.js. It needs no dependencies; z (zod) and some helpers are passed in. The bundled sites/gemini.js and sites/notebooklm.js are full examples.

export default ({ z, kit }) => ({
  url: 'https://news.ycombinator.com/',
  tools: {
    top: {
      description: 'Titles and links of the Hacker News front page',
      input: { limit: z.number().optional() },
      run: async (page, { limit = 10 }) => {
        await page.goto('https://news.ycombinator.com/');
        const links = await page.locator('.titleline > a').evaluateAll(as => as.map(a => `${a.innerText}\t${a.href}`));
        return links.slice(0, limit).join('\n');
      },
    },
  },
});

Mark a tool writes: true if it changes something. It then obeys a site's read-only and confirm settings, and clients see it as destructive. Mark it readOnly: true if it only reads, as top could.

For a tool that takes a while, run gets a fourth argument, { progress }. Call progress(done, total, message) now and then, e.g. progress(30, 300, 'Waiting for the reply (30 s)'). Clients that show progress get it, and for the rest it does nothing.

Then run npx -y anysite-mcp login hn (or just use site: "hn"), and the agent gets hn_top. A pack in ~/.anysite-mcp/sites with the same name as a bundled one replaces it. Pull requests with new packs are welcome.

How it works

• Each site's Chrome runs with its own --user-data-dir in ~/.anysite-mcp/profiles/<site> and a debugging port it picks itself. The server reconnects to it if it is still open, so tabs survive restarts.
• Sign-in uses a plain Chrome with no debugging port, because Google and others block sign-in in automated browsers.
• Elements are found with Playwright's role, label and text locators, never CSS selectors. When a click is blocked by an overlay, the element is operated from the keyboard instead.

Permissions and safety

The agent acts as you on the sites you sign in to. Each site can be limited, and only you can change the limits (the agent has no tool for it):

npx -y anysite-mcp permissions mail.google.com --read-only     # no sending, deleting, posting or uploading
npx -y anysite-mcp permissions github.com --confirm            # ask me first
npx -y anysite-mcp permissions github.com --allow github.com,githubusercontent.com
npx -y anysite-mcp permissions github.com                      # show the current settings

[!WARNING] These are guard rails, not a sandbox: a button that saves something through script, with a plain label like "Go", can get past them, so still only sign in where you're happy for the agent to act. Settings live in ~/.anysite-mcp/settings.json. Profiles hold your cookies: never share or commit ~/.anysite-mcp/profiles.

Security: to report a vulnerability, see SECURITY.md. Please don't open a public issue.

FAQ

Development

npm install
npm test                 # unit tests, then a headless self-check against a local test shop
npm run test:unit        # unit tests only (no Chrome needed)
npm run lint             # ESLint and a TypeScript check of all the code
npm run check:versions   # package.json, package-lock.json, server.json and CHANGELOG.md agree
npm run check:licenses   # every runtime dependency has an allowed license
npm run audit            # npm audit of runtime dependencies; fails on high or critical
npm run test:pack        # pack, install and start the package as npm would publish it
npm run test:inspector   # check the MCP contract with a second client, the MCP Inspector CLI

Releases are published to npm from GitHub Actions with provenance, after CI passes on the tag. Every change to main needs a code-owner review. See CONTRIBUTING.md for the details; new site packs are especially welcome.