WebMCP
A proposal and code for websites to support client side LLMs
WebMCP allows websites to share tools, resources, prompts, etc. to LLMs. In other words, WebMCP allows a website to be an MCP server. No sharing API Keys. Use any model you want.
Here's a simple website I built that is WebMCP-enabled
It comes in the form of a widget that a website owner can put on their site and expose tools to give client-side LLMs what they need to provide a great UX for the user or agent.
The look, feel, how it's used, and security are all absolutely open for contribution / constructive criticism. MCP Clients directly building WebMCP functionality seems like an ideal outcome.
An end-user can connect to any number of websites at a time - and tools are "scoped" (by name) based on the domain to simplify organization.
Super Quick Demo (20 seconds, Sound on 🔊)
https://github.com/user-attachments/assets/61229470-1242-401e-a7d9-c0d762d7b519
Getting started (using your LLM with websites using WebMCP)
Installation
Just specify your MCP client (claude, cursor, cline, windsurf, or a path to json)
npx -y @jason.today/webmcp@latest --config claude
If you're interested in setting it up manually, use the command npx -y @jason.today/webmcp@latest --mcp.
Auto-install was inspired by Smithery, but their code is AGPL so I wrote this myself. If it doesn't work for you or you don't see your mcp client, please file an issue.
Using WebMCP
When you're ready to connect to a website, you can ask your model to generate you an mcp token.
Copy the token and paste it to the website's input. As soon as the website registers with it, it's thrown away and cannot be used for subsequent registrations or anything else. The website will receive its own session token for making requests.
If you'd rather your model / service never see the token, you can manually execute npx @jason.today/webmcp --new instead.
Some MCP clients, including Claude Desktop, need to be restarted to get access to new tools. (at least at time of writing)
To disconnect, you can close the browser tab, click "disconnect", or shut down the server with npx @jason.today/webmcp -q.
All configuration files are stored in ~/.webmcp directory.
Getting started (adding WebMCP to your website)
To use WebMCP, simply include webmcp.js on your page (via src or directly):
<script src="webmcp.js"></script>
The WebMCP widget will automatically initialize and appear in the bottom right corner of your page. Clicking on it will ask for a webmcp token which the end-user will generate.
Full Demo (3 minutes)
https://github.com/user-attachments/assets/43ad160a-846d-48ad-9af9-f6d537e78473
More Info About How It Works
The bridge between the MCP client and the website is a localhost-only (not accessible to requests outside your computer) websocket server. Because it is configured to allow requests from your local web browser, authentication / token exchange is required, in case you visit a website attempting to abuse this.
Ideally the web browser itself would have an explicit permission for this, like webcam or microphone use.
- The MCP client connects to the
/mcppath using the server token from.env(auto-generated) - The server generates a registration token (instigated via the built-in mcp tool by a model or the
--newcommand) - Web clients connect to the
/registerendpoint with this token and its domain. - Web pages connect to their assigned channel based on their domain.
- When an LLM wants to use a tool / resource / prompt, the request flows from:
- MCP Client → MCP Server → WebSocket Server → Web Page with the tool / resource / prompt
- (similar for requesting a list of tools / resources / prompts)
- The web page performs the request (e.g. call tool) and sends the result back through the same path
- Multiple web pages can be connected simultaneously, each with their own set of tools and tokens
- The MCP client sees all tools as a unified list, with channel prefixes to avoid name collisions
sequenceDiagram
participant User
participant MCP as MCP Client
participant Server as MCP Server
participant WS as WebSocket Server
participant Web as Website
%% Initial connection
MCP->>Server: Connect to /mcp with internal server token
%% Website registration token
User->>MCP: Request registration token
MCP->>Server: Request registration token
Server-->>MCP: Return registration token
MCP-->>User: Display registration token
%% Website registration
User->>Web: Paste registration token
Web->>WS: Connect to /register with token & domain (registration token deleted)
WS-->>Web: Assign channel & session token
Web->>WS: Connect to assigned channel
%% Tool interaction
MCP->>Server: Request tools list
Server->>WS: Forward request
WS->>Web: Request tools
Web-->>WS: Return tools list
WS-->>Server: Forward tools list
Server-->>MCP: Return tools list
%% Tool execution
MCP->>Server: Tool request
Server->>WS: Forward request
WS->>Web: Execute tool
Web-->>WS: Return result
WS-->>Server: Forward result
Server-->>MCP: Return result
%% Disconnection
User->>Web: Disconnect
Web->>WS: Close connection
Security
This is a super early project. I'm very interested in hardening security to prevent malicious extensions etc. from being able to perform prompt injection attacks and similar. If you have constructive ideas, please reach out or open an issue.
Built in tools
- Token generator (for connecting to WebMCP websites)
- MCP Tool Definer (to simplify building the schema of a tool for use with MCP)
- You can ask for the javascript (if relevant) in a follow-up message for use with WebMCP
Docker
There is a Dockerfile specifically for Smithery deployment.
If you'd like to use docker to run the websocket server, I've added a docker-compose.yml for demonstration purposes.
If --docker is provided to the mcp client config alongside --mcp, it will assume the server is running. This will allow you to dockerize the main process (websocket server), and your mcp client will connect to your docker container via websocket. Similarly, websites will communicate with your docker container.