Jason McGhee's WebMCP: Turning Websites Into AI Tool Servers

Table of content

Jason McGhee builds tools that make AI more useful in everyday workflows. His WebMCP project lets any website act as an MCP server, allowing AI agents to interact with web services directly in the browser. The idea caught enough attention that W3C is now incubating it as a web standard.

Background

Co-founder and CTO at Writ, a collaborative data platform
Previously co-founded Cursor (data catalog, acquired by DataRobot in 2019)
Former Principal ML Engineer at DataRobot, Sr. Software Engineer at Pandora
Based in San Francisco, builds open source tools focused on local-first AI

GitHub | Twitter | Site

WebMCP: Websites as AI Tool Providers

The core problem WebMCP solves: connecting AI agents to web services typically requires downloading separate MCP servers, managing API keys, or automating browsers with tools like Playwright. McGhee saw a simpler path.

“If there’s a standard instead of needing to download the Jira MCP server, you just visit their website and all the tools are described and usable from there.”

WebMCP works through a browser widget that website owners add to their sites. Users connect their AI assistants using one-time tokens. No API keys change hands. Everything runs client-side.

// Website adds the widget
<script src="webmcp.js"></script>

// User connects via their MCP client
npx -y @jason.today/webmcp@latest --config claude

The architecture uses a localhost WebSocket server:

Component	Role
MCP Client	Connects to local server with token
WebSocket Server	Routes requests between client and websites
Website Widget	Exposes tools, handles authentication

Tools are domain-scoped to prevent naming conflicts. Users can connect to multiple WebMCP-enabled sites simultaneously.

W3C Standardization

McGhee’s proposal gained traction. W3C’s Web Machine Learning Community Group is now developing WebMCP as a browser-native feature.

“WebMCP is being incubated in W3C / webmachinelearning, so highly recommend checking that out as it’s what will turn into WebMCP being in your browser.”

If standardization succeeds, WebMCP becomes a native browser capability. Similar trajectory to Service Workers or Web Workers.

Other Projects

McGhee builds tools that extend what you can do with AI locally:

Project	Description
rem	Records everything on your Mac screen, stores locally, makes it searchable
claude-debugs-for-you	MCP server + VS Code extension for AI-assisted debugging
plock	Use an LLM from anywhere you can type
portable-hnsw	Vector search index served from CDN, runs in browser

The pattern across his work: bring AI capabilities closer to where you already work, keep data local, avoid unnecessary server round-trips.

Key Takeaways

Principle	Implementation
Skip the API key dance	WebMCP uses one-time tokens, no credentials shared
Browser as sandbox	Client-side execution provides security boundary
Build for standards	Proposals that solve real problems attract standardization
Local-first AI	rem, plock, portable-hnsw all process data on device

Links

Next: Simon Willison’s LLM Workflow