uisandbox-mcp
Enables loading a web app build into a live sandbox, adjusting design tokens like colors, typography, and spacing with 1:1 visual verification, and exporting the changes as CSS, JSON, patches, or platform tokens.
README
<p align="center"> <a href="https://uisandbox.org"><img src="brand/marketing/site-hero-light-1440.png" alt="UISandbox — Play with your real app's design, live" width="880"></a> </p>
<h1 align="center">UISandbox</h1>
<p align="center"><b>Play with your real app's design, live.</b><br> Restyle your app without rebuilding it — drop the build, turn the knobs, watch it follow 1:1, export the patch.<br> <sub>Test your design on the real thing.</sub></p>
<p align="center"> <a href="https://uisandbox.org">uisandbox.org</a> · <a href="https://www.npmjs.com/package/uisandbox-mcp">npm: uisandbox-mcp</a> · <a href="https://registry.modelcontextprotocol.io/v0/servers?search=uisandbox">MCP registry</a> · <a href="LICENSE">MIT</a> </p>
<p align="center"> <img alt="MIT" src="https://img.shields.io/badge/license-MIT-black"> <img alt="npm" src="https://img.shields.io/npm/v/uisandbox-mcp?label=uisandbox-mcp&color=black"> <img alt="nothing leaves your tab" src="https://img.shields.io/badge/nothing%20leaves-your%20tab-black"> <img alt="1:1 measured" src="https://img.shields.io/badge/1%3A1-measured%2C%20not%20promised-black"> </p>
npx uisandbox-mcp open ./dist # the sandbox opens in your browser with your build — no agent needed
Or drop a zip on uisandbox.org — no build at hand? Try a sample under the drop zone (the Bootstrap docs, a VitePress site, an admin dashboard). Or, in Claude Code: /plugin marketplace add AlexanderKaan/uisandbox → /plugin install uisandbox@uisandbox and say "open this app in UISandbox".
<p align="center"><img src="brand/marketing/stage-acme-brand-crimson.png" alt="A real build in the sandbox: the Brand knob turned to crimson, the page following, five values moved" width="880"></p>
What it does
- Bring the build — three ways in: a zip of
dist/,build/,out/(or a whole repo with the build inside); a folder (a build renders, a source folder is read for the knob stand); a public GitHub or GitLab repo URL (fetched as a zip through a same-origin route — the one thing that leaves the tab, seenotes/security.md). While it works, the door shows the stages with their numbers. Otherwise nothing leaves the tab: files are read in the page and served to the frame by a service worker on this origin. - See it 1:1. Every colour, radius, font, size, spacing, line-height, letter-spacing, weight, border-width, duration, gradient angle and shadow literal in your CSS becomes a variable holding the very same value, so the page renders exactly as it was — runtime styles too (
style=""set by JS,<style>your framework appends,insertRule/replaceSyncfrom styled-components, Emotion, Lit, Ant's cssinjs; CDN stylesheets through a same-origin proxy; nested same-origin frames such as Storybook's). - Turn the knobs. Brand, and the colour families your CSS actually contains (secondary, accent, the status set, a palette row for what is neither); page background; grey tint and border tone; display and body font (yours listed first, then alternatives by character); dials for text size, line height, letter spacing, weight, spacing, radius, border width, elevation, motion and gradient angle — every dial with ×1 = as in your code at its centre; global hue, saturation and contrast that reach every colour, chart palettes and CSS-drawn icons included; your dark mode, switched on your own hooks.
- Export. Your values as CSS / JSON / a patch list, your files patched in place,
--k-*tokens for web (CSS, Tailwind, shadcn), Swift constants + an asset catalog for iOS,colors.xml+ Kotlin for Android.
Honest by construction
- "1:1" is measured, not felt. Check 1:1 loads the untouched build and the tokenised build side by side and diffs the computed styles of every element (18 properties, shadow roots and nested frames included). Zero differences or it says what differs.
- A reach meter in the stage foot says how much of what you see the knobs touch — painted colours, families, sizes, radii — and what lies outside (images, canvas, video).
- It refuses what it cannot show. iOS/Android projects, WordPress themes and source without a build get a clear message at the door; a page that asks for files the archive does not hold says so ("that usually means source, not the built output").
- A hostile archive cannot navigate the tool away, register its own service worker or unregister ours, or smuggle a zip-slip path back out of an export. What it can do — reach the tool's DOM, because the frames are same-origin by design — reaches nothing worth having: no server, no credentials, no state. Read
notes/security.md; deploy on an origin of its own.
Run it
pnpm install
pnpm dev # http://localhost:5190
pnpm test # vitest + node:test (audit engine)
pnpm build # dist/ — static; public/_headers rides along
pnpm holdouts # every fixture zip through the real app in headless Chromium (see below)
http://localhost:5190/?load=<zip-url> loads a zip by URL (same-origin or CORS) — the door an agent uses.
Hold it to account
pnpm holdouts runs every archive in fixtures/ (real builds from public repos — gitignored, notes/decisions.md says where each came from) through the app: load, Check 1:1, reach, and a host check (same origin, one worker). Verdicts are held against scripts/holdouts.expect.json; a fixture expected ok that is not fails the run. --only <name>, --record, --base <url> --fixtures <url> for a production build or the live site.
How it works
src/
sandbox/
rewrite.ts their CSS/HTML → literals replaced by var(--us-vN); byte-preserving
table.ts the substitution sheet: one entry per (kind, value), with sites
mapping.ts knob → their values, RELATIVE to the baseline (identity at rest); families
baseline.ts the knobs on the stand of THEIR code (audit + the sheet + the rendered page)
project.ts root and deploy-base detection · screens · raw + rewritten files · guard + hook
host.ts owns the sandboxes, answers the service worker; live vars injected into HTML
verify.ts the 1:1 check: raw vs identity, computed styles per element
coverage.ts the reach meter
live.ts MutationObserver: runtime style="" and <style> get the same rewrite
scheme.ts their dark mode: hooks found in their CSS, switched
public/sw.js serves a sandbox document's URLs from the page; /__ext/ proxies CDN CSS
tokens/ panel/ state/ export/ audit/ the knob engine, panel, undo/hash, exporters, source audit
notes/ decisions (numbered), traps, lessons, security, roadmap
MCP
npx -y uisandbox-mcp (npm) — or pnpm mcp from a clone — runs the same engine as an MCP server (stdio): load a zip by URL or path, set knobs, export any format, verify the 1:1 check and screenshot in headless Chromium against the real app. See mcp/README.md for the tool list and the Claude/Cursor config. Claude Code plugin (skill + server in one): /plugin marketplace add AlexanderKaan/uisandbox then /plugin install uisandbox@uisandbox — gives /uisandbox ("open this app in a sandbox", "try brand #e11d48") and the uisandbox MCP server. The skill alone is skills/uisandbox/SKILL.md (copy into .claude/skills/). Codex / Cursor / others: add the MCP server (npx -y uisandbox-mcp); the server sends its own instructions and prompts.
Deploy
One origin, one Cloudflare Worker (worker/index.mjs, wrangler.jsonc): static assets from dist/, the SPA fallback for ?load=, http→https and www→apex redirects, and the /__repo/ route that fetches a public GitHub zip for "Connect a repo" (nothing stored; same-origin callers only; a rate-limit rule in the zone). public/_headers keeps sw.js uncached and sets the security headers. Put nothing else on the origin — the sandboxed frames are same-origin by design (notes/security.md).
Roadmap
notes/roadmap.md: ship (uisandbox.org) · the intake as a front door · being found (SEO, llms.txt, analytics) · the MCP server · after launch.
Made with ♥ by Alexander Kaan at Pageminds · MIT, free forever · the decisions that shaped it, numbered: notes/decisions.md
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.