website-mcp — Website to Deterministic MCP Tool

Explore a site once. Replay it forever, free.
Claude Code writes and tests a script for a site task, saves it as an MCP tool, and from then on it runs with no model in the loop.

You type /website-mcp:mint <task> on <site>. Claude explores the site, writes a script, proves it against test cases, gets it reviewed, and saves it as a tool. After that the tool is a plain file: it runs instantly, costs nothing, and gives the same answer shape every time.

The map

Related MCP server: agentic-browser-mcp

Quick start

Needs Node 22+ and Claude Code. Chrome is the supported browser for page tools.

npm install && npm run build
claude plugin marketplace add .
claude plugin install website-mcp@website-mcp

1. Mint a tool — in Claude Code run /website-mcp:mint <what to read> on <site url>; pick "browser" for public sites.

2. Run it anywhere — node dist/cli.js run <name> --input '{"q":"..."}' from the checkout (or npm link once, then plain website-mcp run …), or call it as an MCP tool. Below, website-mcp means either form.

The local marketplace above loads the checkout in place. A marketplace added from a GitHub URL has no dist/: set WEBSITE_MCP_ALLOW_NPX=1 (runs the published npm package) or WEBSITE_MCP_ROOT (path to a built checkout). Details in plugin/README.md.

Browser tools work now. Page tools need one more setup, below. That's it for the basics; everything below is detail.

Demo

Output of the test suite (test/browser) running a saved tool against the repo's local fixture site (test/fixtures/site/), no model involved. It is not a command you can paste from a fresh checkout: the fixture server and origin are set up by the tests.

$ website-mcp test fixture_search
case          result  ms     error
tag_alpha     PASS    174
tag_beta      PASS    162
tag_gamma     PASS    163
no_match      PASS    161
blank_query   PASS    116
empty_string  PASS    1

6/6 passed (100%); bar: >=3 cases and >=70% -> OK

$ website-mcp run fixture_search --input '{"q":"alpha"}'
{
  "items": [
    "Fixture item 01",
    "Fixture item 06",
    "Fixture item 11",
    "Fixture item 16",
    "Fixture item 21"
  ]
}

Page tools setup (your own Chrome)

Page tools run in your real, signed-in tab, so they need the extension and a native-messaging bridge.

1. Load the extension — open chrome://extensions, turn on Developer mode, choose "Load unpacked" and pick the extension/ folder.

2. Allow User Scripts — on the extension's details page, turn on Allow User Scripts. Without it the extension cannot run tool code.

3. Install the bridge — node dist/cli.js bridge install --browser chrome (or website-mcp bridge install … after npm link), then restart Chrome.

4. Check — website-mcp bridge status should report the host installed and the extension connected.

5. Grant and approve — in the extension popup, click Allow this site for each site, and Approve for each tool version. Changing a tool's code asks again.

Draft, test, publish

Minting goes through a draft folder, never straight into tools/:

1. save_tool writes the draft tool.json and tool.mjs.

2. save_cases writes at least 3 cases, including an empty and an invalid_input one.

3. test_tool runs them through the normal runtime, no model.

4. publish_tool re-checks a staged copy and publishes only if the bar is met: at least 70% of cases pass, valid schemas, no secret-shaped literals, no URL outside the allow-list, and a declared effect.

The previous version moves to .history/<n>/. Commands: serve, list, run, test [--draft], review, publish, bridge install|status|uninstall.

Safety

A page tool is model-written code running in your signed-in tab, so read docs/safety.md. In short: the static scan is not a sandbox; what protects you is the per-site grant, the per-version approval bound to the executed code, the explore window, and, for writes, you pressing the final button. The page also covers site terms and what is and is not a boundary.

Heads up: a tool you mint acts as you. Check a site's terms before pointing a tool at it.

Status

Phases 1 to 8 are built (see CHANGELOG.md). Not yet: a Chrome Web Store listing, a local-model minter, auto-repair. Prior art and what was and was not borrowed: docs/prior-art.md.

License

MIT, see LICENSE.