mcpbuild

Chat to a tool.
Live in your MCP client.

Ask Claude or ChatGPT to build a widget inside MCPBuild. Each project becomes a connectable MCP server — same widget, both clients, with live data and version history.

free forever tier· no credit card· ships in 60 seconds
mcpbuild.app / new
live
renders inline in
Claude·ChatGPT
connect the MCP from any client
Cursor·Zed·any MCP client
How it works

Three steps, end-to-end.
No infra, no glue code, no marketplaces.

01
build 30 seconds

Chat your way to a widget.

Describe what you want in plain English. Claude or ChatGPT writes the React, fetches real data, and renders it inline — right inside MCPBuild.

> Build me a quarterly revenue chart from analytics.acme.com/api, grouped by product line.
mcpbuild.app / new
live
02
connect 1 click

Every project is an MCP endpoint.

Each project comes with its own MCP server URL. Add it once to Claude, Cursor, or any MCP client — your widget is available as a callable tool.

~/.claude/mcp.json copy
{
  "mcpbuild": {
    "url": "https://mcpbuild.app/mcp/quarterly-revenue"
  }
}
mcpbuild.app / project / connect
live
03
share shareable link

Public widgets, no login required.

Flip the toggle to make a project public. Share a URL — anyone can use the widget directly in ChatGPT or open it as a hosted page.

mcpbuild.app/p/quarterly-revenue
Copy
chatgpt.com
live
What's in the box

Everything you need to ship a tool — nothing you don't.

live data

Real APIs, not fakes.

Widgets call your endpoints with credentials scoped to the project, injected server-side. No copy-pasted JSON, no demo data, no leaked keys.

versioning

Every save is a checkpoint.

Remix a widget in chat and it saves a new version. Roll back to any earlier version in one click — full history is kept.

two clients, one widget

Claude and ChatGPT, in parity.

Build it in either chat client. Connect it as MCP. Same component, same behavior — no separate forks.

mcp-native

Each project is an endpoint.

Projects come pre-wired as MCP servers. Add one URL to your client config — every tool inside becomes callable.

public sharing

A link is enough.

Flip a project public. Drop the URL into a ChatGPT conversation — your widget renders inline. No installer, no auth wall.

private by default

Private until you say otherwise.

Projects start private and stay that way. Per-project, KMS-encrypted secrets and a fetch allowlist on the Pro plan.

What people build

From dashboards to one-off calculators.

$1.24M+12.4%
Quarterly Revenue
finance
78
healthy · 412 accounts
Customer Health Score
success
Cchenshipped auth migration
Ppriyablocked on stripe webhook
DdevPR #482 review needed
Sprint Standup Brief
engineering
api
edge
db
auth
search
billing
On-call Status Board
platform
seats × 24$456
SSO add-on$120
annual discount-$144
total / mo$432
Pricing Calculator
sales
P1edge worker timeout in us-east
P2stripe webhook lag (3m)
P3logs slow on /metrics
Active Incidents
support
Pricing

Honest pricing. Real free tier.

No seat math, no AI-token meters, no "contact us" wall. Start free, upgrade when you outgrow one project — or buy it once and never think about it again.

Free
$0
free, forever
Build, remix, and connect. Perfect for personal tools and weekend projects.
  • Unlimited widget generation & remix
  • The Inbox for your renders
  • 1 project you can connect as an MCP
  • Static data
most popular
Pro
$8
per month
For builders shipping real, reusable tools — and sharing them.
  • Everything in Free
  • Unlimited projects
  • Version history & restore
  • Live data — bring your own API keys
  • Publish public MCP endpoints + galleries
  • Usage analytics
  • Export any widget or project (.zip)
no credit card to start · cancel anytime · secure checkout via stripe
FAQ

Questions, answered straight.

MCP — the Model Context Protocol — is an open standard for connecting AI assistants to tools and data. MCPBuild gives every project its own MCP server URL, so you can plug it into Claude, ChatGPT, Cursor, or any other MCP-aware client. You don't need to understand the protocol — just paste the URL into Claude or ChatGPT's connector settings and your tools are there.
Yes, that's the design target. You describe what you want in a chat with Claude or ChatGPT ("a dashboard that shows me orders with a Cancel button per row") — the AI writes the React for you, MCPBuild hosts it, and the same widget renders inline in Claude, ChatGPT, and on your project's web page. No code, no deploy step. You get a shareable link and an MCP URL the same minute.
Connect both — they do different jobs. The workshop MCP (mcpbuild.app/mcp) is where you BUILD: it gives you the dynamic_widget, project, and remix_widget tools so the AI can author and save tools for you. The project MCP (mcpbuild.app/mcp/<your-project>) is where the tools you built actually live and get CALLED by the model. Rule of thumb: connect the workshop in your day-to-day Claude/ChatGPT so you can prototype anywhere; connect a specific project endpoint when you (or someone you're demoing to) wants to USE that project's tools without seeing the build surface.
Three options. (1) Send them the project's public share link — they get the widget rendered in a webpage, no connector needed. (2) Have them connect your project's MCP URL in their own Claude/ChatGPT and just chat naturally. To make natural chat actually trigger your tools, add a few “demo prompts” on the Triggers tab — phrases a real user would say, like “cancel order A-100” or “show me this week's pipeline.” Those phrases are appended to the tool's description so the model picks the right tool without you having to coach the audience. (3) On the project page, click any of the saved demo-prompt pills to copy a working prompt to your clipboard.
A trigger phrase is a short sentence a real user would say to invoke your tool — for example, “cancel order A-100”, “refund the customer named Alice”, or “show me this week's pipeline.” You add them on each tool's Triggers tab. Two things happen: (1) they're appended to that tool's MCP description so Claude and ChatGPT pick the right tool when someone phrases the request naturally — you no longer need to tell the audience the tool's name; (2) each phrase becomes a one-click “Try this prompt” pill on the project page so anyone giving a demo can copy a working prompt to their clipboard. Rule of thumb: 3–5 phrases per tool, written the way a non-technical user would actually speak.
It's in your Inbox. By default, prompts like “use dynamic_widget for a chart…” drop the widget into a 24-hour ephemeral Inbox so you can iterate without polluting your projects. To save it into a named project in the same step, tell the AI which project — e.g. “use dynamic_widget with projectName: 'Orders' …” — and it auto-saves the widget as a callable tool on /mcp/orders, no second step required. If a widget is already in the Inbox, open it and click Save to project to keep it.
It used to fail silently because the page didn't surface widget runtime errors. We fixed that — broken widgets now print their actual JavaScript error inline in the iframe instead of going blank. If you see a stack trace there, copy it back into the same chat and ask the AI to fix it (e.g., “the widget threw `Cannot read properties of undefined (reading 'map')` — fix it”). Usually it's a one-shot remix_widget call.
No. The AI writes the React for you. You refine it by chat ("make the chart darker", "sort by date", "add a column for the customer email") and never touch code — but every widget is plain TSX you can read, edit, or download if you want to.
Artifacts and canvases live inside one conversation — they vanish when you start a new chat. MCPBuild widgets persist as projects, version every change, get their own MCP endpoint, and work across both chat clients — same widget, no rewrite. You can also share them via a public URL or invite teammates without anyone having to copy code around.
Widgets and projects live on AWS, encrypted at rest. If a widget needs to fetch from a real API (say, your CRM), project secrets are KMS-encrypted, scoped to that one project's MCP endpoint, and injected server-side at fetch time — the model never sees them in plaintext. For demos, prefer pasting fake data into a widget; for real demos, the secrets+allowlist flow keeps your keys off the screen.
Claude (Sonnet, Opus) and ChatGPT (GPT-4-class models). MCPBuild never runs inference itself — the model on the other end of the chat writes the React, and we validate, bundle, and host it. That means we pass through to whatever account you already pay for — no extra LLM bill.
Not today — it's a hosted service. Your widgets are portable, though: download any widget or a whole project as a self-contained .zip and re-render it anywhere.

Stop pasting screenshots
into your AI chats.

Your AI already knows what to build. MCPBuild gives it a place to put it — and you a way to use it again tomorrow.

1 project free · no credit card · ships in 60 seconds