Websites are starting to describe their own actions for AI agents. The standard for this is WebMCP: a page registers tools such as add_to_cart or place_order with document.modelContext.registerTool, or marks up a form with toolname. Chrome has it in an origin trial, and shops have started to turn it on.
Usher connects your AI agent to those tools in the browser tab you already have open and are logged into. It is a Chrome/Edge extension plus a small local MCP server. Your agent (Cursor, Claude Desktop, Claude Code, or any MCP client) asks Usher what the page can do, and calls it. When a tool can pay, send, book or delete something, Usher shows an and waits for you.
PlugMCP Browser (Usher) forwards named page tools to Cursor, Claude, and other MCP clients. Consent stays in the page. It is not a hosted gateway server.
—
Allow / Deny bar inside the page
Usher is not a new browser, not a click-and-screenshot agent, and not a new protocol. It is the door between your agent and the page, with you standing at it.
What it does
Finds the page's tools. Three kinds: WebMCP tools the site registers, declarative WebMCP forms, and tools you saved yourself.
Asks before risky actions. Paying, ordering, sending, booking, deleting and submitting forms wait for your Allow. No answer in 90 seconds means no.
Works on silent pages too. On a page with no tools, your agent can propose tools drafted from the page's forms. Only you can save them.
Stays in your tab. By default it shares only the tab you are looking at. It never fills password or card fields.
Shows what happened. A small note in the corner ("Cursor ran set_address"), a tool count on the toolbar badge, and recent activity in the popup.
Save tools for a silent page
A saved tool asks before it submits
The popup
Quickstart (about 2 minutes)
You need Node.js 20+ and Chrome or Edge. Full step-by-step instructions for Windows and macOS are in the easy guide.
1. Build it
git clone https://github.com/demo55oo/usher
cd usher
npm install
npm run build
2. Load the extension. Open chrome://extensions (or edge://extensions), turn on Developer mode, click Load unpacked, and choose the folder packages/extension/dist.
3. Add Usher to your agent. For Cursor, put this in ~/.cursor/mcp.json (or .cursor/mcp.json in a project). Use the full path to your copy:
Claude Code: claude mcp add usher -- node /full/path/to/usher/packages/server/dist/usher-mcp.js. Other agents: see docs/agents.md.
4. Try the demo shop. Run npm run demo, open http://localhost:5178/coffee/ in the browser with Usher, and ask your agent:
Order two flat whites to 12 Elm Street, Springfield 49007 with the code CORNER10.
The agent fills the cart, address and code by itself. When it tries place_order, Usher asks you first.
No agent at hand? npm run demo:agent runs a scripted agent against your browser, and npm run demo:agent -- --self runs the whole thing in a headless browser.
How it works
Your agent ── MCP (stdio) ──▶ usher-mcp ── WebSocket 127.0.0.1:47800–47809 ──▶ Usher extension
(Cursor, Claude…) (local server) (only extension origins allowed) │
▼
your real, logged-in tab: WebMCP tools │ declarative forms │ saved tools
│
risky? ──▶ Allow / Deny bar in the page
The server is a normal MCP stdio server. It opens a WebSocket on the first free port from 47800 to 47809 and accepts connections only from browser extensions. The extension connects to every Usher server it finds, so several agents can use the same browser. Inside the tab, a small script sees the page's WebMCP tools (native, or through Usher's spec-shaped polyfill), and a separate isolated script runs the consent bar where page code cannot reach it. Details: architecture, protocol.
The five tools your agent gets
Tool
What it does
Changes things?
list_tabs
Lists the tabs you share (by default: the one you are looking at) and how many tools each has.
No
list_page_tools
Lists the page's tools with their input schema and gate (auto or ask).
No
call_page_tool
Runs one tool in your tab. Tools with gate ask wait for your Allow.
Yes, with your yes
propose_page_tools
Drafts tools from the forms on a silent page and asks you to save them.
Only if you click Save
read_page
Read-only summary: title, headings, visible text, forms (labels only).
No
Security, in short
Risky calls wait for a trusted click on Allow, in a closed shadow root, after the buttons have been visible for half a second. Page scripts cannot click it, hide it or fake it.
The decision is made in the extension, never in the page's JavaScript world.
Only the tab you are looking at is shared, unless you turn on "Share all tabs".
Usher never fills password, card number, CVC or one-time-code fields, and never reads what you typed.
Page text and tool results are marked as untrusted content for the agent.
The local server listens on 127.0.0.1 only and refuses connections from web pages.
How is this different from MCP-B?MCP-B made WebMCP usable early: @mcp-b/global is a polyfill sites can ship, and its local relay connects agents to pages that embed its script. Usher works on any tab through an extension and adds the consent bar, tab scope and saved tools. Pages that use @mcp-b/global work in Usher too (covered by an end-to-end test with @mcp-b/global 6.0.1).
How is it different from Browser MCP or browser-use?Browser MCP and browser-use drive a page by clicking and typing, which works everywhere. Usher does not click: it calls the tools the site itself offers, and asks you before the risky ones. They fit together; a browser-use fallback is on the roadmap.
And WebMCP Hub? WebMCP Hub collects WebMCP tool definitions for sites. Usher is the runtime in your browser that calls tools and asks you first.
Does it work with agents Paperclip hires? Yes. Usher is a standard MCP stdio server, so any agent that can start one can use it, wherever that agent came from. The person at the browser still decides on every risky action.
Does my browser need WebMCP turned on? No. If the browser has native WebMCP, Usher uses it. If not, Usher brings a polyfill shaped like the current draft spec, so sites that feature-detect document.modelContext register their tools into Usher.
Is it on the Chrome Web Store or npm? Not yet. For v0.1 you load the extension unpacked and run the server from the built file.
Does it see my other tabs, passwords or history? No. It sees the tab you are looking at, only when an agent asks, and it skips password and payment fields.
Roadmap
Voice and remote agents, iframes, Firefox, a browser-use fallback for pages without tools, Chrome Web Store and npm releases, per-tool "always allow", and multiple accounts. See docs/roadmap.md.
Development
npm run typecheck # TypeScript, all packages
npm test # unit tests (vitest)
npm run test:e2e # builds, then runs the real extension in headless Chromium
npm run package # release/usher-extension-<version>.zip and release/usher-mcp.js
PlugMCP Browser (Usher) forwards named page tools to Cursor, Claude, and other MCP clients. Consent stays in the page. It is not a hosted gateway server.
Tools
Tools
Noch keine Tool-Liste. Remote-Server erhalten eine nach dem ersten Health-Check.
FAQ
Fragen
Wie installiere ich PlugMCP Browser?
Kopieren Sie auf dieser Seite die stdio-Konfiguration für Cursor, Claude Desktop, Claude Code oder VS Code. Sie nutzt die Installation des Herausgebers, keinen PlugMCP-Proxy.
In diesem Eintrag ist keine erforderliche Umgebungsvariable aufgeführt.
Welchen Transport verwendet PlugMCP Browser?
PlugMCP Browser läuft lokal über stdio.
Wann wurde PlugMCP Browser zuletzt geprüft?
Zuletzt geprüft: 2026-11-05. PlugMCP Browser (Usher) forwards named page tools to Cursor, Claude, and other MCP clients. Consent stays in the page. It is not a hosted gateway server.