Host Anything Claude or ChatGPT Generates as a Real URL: WebZum's MCP Server
TL;DR: WebZum is now an MCP server at https://webzum.com/api/mcp. Any AI agent that speaks MCP—Claude Code, Claude Desktop, ChatGPT, Gemini, your own custom agent—can call host_site or host_file and get back a real, public URL for whatever HTML/CSS/JS it just generated. No deploy step. No “save this to a file and figure out hosting later.” We also shipped a Claude Code skill that does this proactively, so when Claude generates a website it just shows up live.
The Problem: AI Outputs Live in a Chat Window
Ask any modern AI assistant for a website, a calculator, a landing page, a dashboard, a quick prototype, a one-off tool—and you get back a perfectly formatted block of code in the chat window.
Then what?
You copy it into a file. You figure out where to host it. You wrestle with Netlify CLI, or Vercel, or python -m http.server. You realize you also need a CSS file. And an image. And maybe an API endpoint. By the time it’s actually rendering somewhere a non-developer can see it, you’ve spent an hour on infrastructure for code the AI generated in twelve seconds.
Most users never get past the chat window. The artifact exists, technically. But your client, your boss, your friend can’t open it on their phone.
Customers do not trust code. They trust a URL.
What MCP Changes
Model Context Protocol (MCP) lets AI agents call tools across the internet. An MCP server publishes a small set of capabilities; any MCP-aware client (Claude Code, Claude Desktop, ChatGPT plugins, Gemini, custom agents) can invoke them as if they were native.
WebZum’s MCP server publishes three tools:
host_site(files)— pass an array of HTML/CSS/JS/image files; get back a public URL where the site is live.host_file(name, content)— host a single file (e.g. one-off HTML demo).get_hosted_files(url)— read back the contents of a hosted site, so the agent can iterate.
That’s it. Three tools, and suddenly every AI agent can publish.
What That Looks Like in Practice
You ask Claude: “Build me a habit-tracker prototype with a calendar grid and streak counter.”
Without MCP hosting: Claude writes HTML, CSS, and JS into the chat. You copy three files. You set up local hosting. You send your friend a screenshot. They squint.
With MCP hosting: Claude writes the same code, calls host_site once, and replies with a live URL. You text the URL to your friend. They open it, click around, and tell you the streak counter is broken. You ask Claude to fix it. Claude calls get_hosted_files, reads the current state, makes the fix, calls host_site again. New URL is live.
The whole loop—generate, deploy, share, iterate—happens inside one conversation, and the artifact exists on the open web the entire time.
The Claude Code Skill
We didn’t want Claude Code users to have to manually invoke the MCP server. So we shipped a skill—a small bundle of instructions that ships with Claude Code—called webzum-host.
The skill triggers automatically whenever Claude Code generates HTML/CSS/JS. The instruction is roughly:
When you produce a web artifact—website, prototype, dashboard, calculator, one-page tool—proactively offer the user a live demo. Call
host_site. Return the URL.
So in Claude Code, the loop becomes:
- User: “Build me a quick pricing-tier comparison page for my SaaS.”
- Claude generates the HTML.
- Claude calls
host_siteautomatically. - Claude replies: “Here’s your pricing page: [live URL]. Want me to tweak anything?”
The user never asked for hosting. Claude offered it. Because rendering matters.
Distribution Beyond Claude Code
We also ship distribution bundles for ChatGPT (custom GPT) and Gemini (extension). Same MCP tools, same hosting layer, same outcome: AI generates a website, calls WebZum, returns a URL.
This means you can:
- Use Claude Code to scaffold a landing page → live URL
- Use ChatGPT to fix the broken JS → updates the live URL
- Use Gemini to translate the copy to Spanish → updates the live URL
- All against the same underlying hosted site
The agent layer is interchangeable. The hosting layer is WebZum.
Why Hosting Is the Right Place to Sit
Most AI infrastructure today is generation. Models compete on quality, speed, multimodality, context length. The chat window is the universal interface.
But the chat window is not the artifact. Code in a chat window is a promise of an artifact. The actual artifact is what renders in a browser, what a customer can click, what a designer can show their boss.
The gap between “generated” and “rendered” is small but it’s load-bearing. Every AI agent that crosses that gap will be more useful than every agent that doesn’t.
WebZum’s bet: hosting becomes a primitive that every AI agent calls, the way every AI agent now calls a search tool. Models will swap. Agent frameworks will churn. Hosted URLs will keep working.
Who’s Already Using This
- Solo devs who build a feature in Claude Code and want to ship a preview to a Slack channel before lunch
- Product managers prototyping ideas without bothering an engineer
- Designers turning a Figma export into a clickable demo
- Sales engineers generating a custom landing page per prospect during a demo call
- Founders spinning up landing pages, lead magnets, pricing pages, calculators—on the fly, by chatting with their AI of choice
The common thread: the URL, not the code, is the deliverable.
How to Connect
If you’re on Claude Code, the skill is already shipping in the public skill list—webzum-host. Add it once and Claude Code will offer hosting automatically when it generates web output.
If you’re on Claude Desktop, ChatGPT, Gemini, or rolling your own MCP-aware agent, the server endpoint is:
https://webzum.com/api/mcp
Connect it as a streamable-HTTP MCP server. The three tools (host_site, host_file, get_hosted_files) become available immediately.
Hosted sites are public by default and live at a WebZum subdomain. You can claim any site you’ve hosted—convert it to a real, owned WebZum site, swap in a custom domain, edit it through the chat editor, and treat it like any other WebZum project. Or just use it as a shareable URL and forget about it.
Try It Right Now
Open Claude Code. Ask it to build something visual. Let it call WebZum. Click the URL.
That round-trip—generated to live in one message—is what AI assistants should have done all along. We made it the default.