Skip to main content
Glama

The problem

You ask an AI to "make the hero section look better" and it dumps 200 lines of arbitrary CSS, overwrites your design tokens, ignores your existing component library, and produces something that looks like every other AI-generated site.

Atelier fixes that.

Related MCP server: ui-architect-mcp

What this actually does

Atelier MCP sits between your AI coding assistant and your frontend project as an MCP server. It exposes 10 tools that give the AI structured context about your project before it touches anything:

Your AI assistant
      │
      ▼
  ┌──────────────┐     inspect → understand tokens, components, conventions
  │  Atelier MCP │     search  → find recipes that match the project's aesthetic
  │   (stdio)    │     plan    → structured changes with rollback capability
  └──────────────┘     apply   → safe changes with automatic backups
      │
      ▼
  Your website project

The AI inspects first, consults a curated recipe catalog, builds a plan, and only then makes changes — with backups so you can undo everything.

One-line install

macOS / Linux:

curl -fsSL https://raw.githubusercontent.com/YOUR_USERNAME/atelier-mcp/main/install.sh | bash

Windows (PowerShell):

irm https://raw.githubusercontent.com/YOUR_USERNAME/atelier-mcp/main/install.ps1 | iex

Both scripts check prerequisites (Node 20.19+ (20.x) or 22.12+, Git), clone the repo, install dependencies, build, and drop the atelier CLI onto your PATH. Done.

Quickstart

# clone it
git clone https://github.com/YOUR_USERNAME/UiMCP.git
cd UiMCP

# install
npm install

# build everything
npm run build

# install into your website project
npx tsx packages/cli/src/cli.ts install --source . /path/to/your/site

# or just run the MCP server directly for debugging
npm run dev:server

Hook it up to Claude Code

claude mcp add atelier-mcp -- npx tsx /path/to/atelier-mcp/packages/server/src/index.ts

That's it. Your AI now has design tools.

Monorepo layout

.
├── packages/
│   ├── server/       →  the MCP server itself (10 tools over stdio)
│   ├── cli/          →  install/uninstall/doctor/update commands
│   ├── recipes/      →  curated design & motion recipe catalog
│   ├── adapters/     →  framework detection (React+Vite, Next.js)
│   └── monitor/      →  Neurolink web dashboard (real-time visualizer)
│
├── install.sh        →  one-line installer (macOS/Linux)
├── install.ps1       →  one-line installer (Windows)
├── LICENSE           →  MIT
├── THIRD_PARTY_LICENSES.md
└── package.json      →  npm workspaces monorepo root

The tools

Tool

What it does

inspect_project

Reads your project — framework, components, design tokens, config files, existing conventions. The AI calls this first.

search_recipes

Searches a curated catalog of motion & UI patterns by purpose, aesthetic, framework, or performance cost.

get_recipe

Returns full implementation code for a recipe — React component, CSS, dependencies, a11y metadata, license info.

prepare_design

Creates a structured plan: design direction, tokens, selected recipes, and exact file changes. Persisted to disk.

preview_changes

Shows what would change before anything is written. Detects conflicts, lists dependencies.

apply_changes

Writes the files. Backs up everything it touches. Returns a record ID for rollback.

rollback_changes

Reverts Atelier changes. Detects if you've edited files since and warns before overwriting.

capture_preview

Screenshots your running site at mobile/tablet/desktop via Playwright.

audit_ui

Runs accessibility checks (axe-core), Lighthouse performance, and contrast validation.

monitor_project

Event log and session state — what tools were called, what changed, project dependency graph.

Recipe catalog

The recipe catalog isn't just "here's some CSS." Each recipe ships with:

  • Full React component code for both Vite and Next.js

  • Accessibility story — keyboard, touch, screen reader, prefers-reduced-motion

  • Performance budget — JS size, CSS size, GPU needs, expected FPS

  • License — know exactly what you're shipping

  • Parameters — every recipe is configurable, not copy-paste-and-pray

Current categories:

Category

Examples

Text reveals

Split text, per-word stagger, line-by-line

Entrances

Staggered grid cascade, fade-scale-in

Buttons

Tactile press feedback, ripple effects

Hover effects

Perspective tilt card, magnetic cursor

Dialogs & menus

Accessible modal with focus trap, slide panels

Route transitions

Cross-fade, slide, scale between pages

Scroll progress

Reading progress bar

Parallax

Restrained depth layers (not the 2014 kind)

Sticky storytelling

Scroll-pinned narrative sections

CLI

atelier install [dir]        # Install Atelier into a project
atelier uninstall [dir]      # Clean removal
atelier doctor [dir]         # Health check
atelier update [dir]         # Update source reference
atelier serve                # Run MCP server directly

Install flags:

--source <path>     Local Atelier source directory
--repo <url>        Clone from a git repo instead
--ref <ref>         Specific branch/tag/commit
--dry-run           See what would happen
--skip-mcp          Don't auto-configure MCP client
--skip-skill        Don't install CLAUDE.md companion file

How it stays safe

  1. Never writes without a plan. The AI must call prepare_design → preview_changes → apply_changes. No yolo file writes.

  2. Backs up everything it touches. Every modified file gets a backup in .atelier/backups/.

  3. Rollback is first-class. rollback_changes restores backups and detects user edits made after Atelier's changes.

  4. Path validation. All file operations are sandboxed to the project directory. No traversal escapes.

  5. Respects what's already there. inspect_project reads existing tokens, conventions, and component patterns so the AI doesn't bulldoze your design system.

Development

npm run typecheck        # type checking across all packages
npm run lint             # ESLint across source and tests
npm test                 # vitest
npm run dev:server       # watch mode on the MCP server

Supported frameworks

  • React + Vite (≥18.0)

  • Next.js (≥14.0, App Router)

More coming. PRs welcome.

A responsive, dark dashboard for watching Atelier in action. Requires Node.js 20.19+ (20.x) or 22.12+ for the development toolchain.

# Start the monitor
npm run dev:monitor -- --project /path/to/your/site

# Open http://localhost:4510
# Use --port 4511 to choose a different port

What you get:

  • Overview — Project metrics, an interactive tool map, recent activity, and plans.

  • Activity — Search and filter the latest 200 events; pause the feed to inspect details.

  • Project — Framework, dependencies, components, design tokens, and configuration files.

  • Design plans — Review saved plans, status, file changes, and related change records.

  • Tool library — Explore all ten tools, their purpose, and recent calls.

  • Snapshot export — Download the current workspace data as JSON.

The monitor connects locally, reconnects automatically, and displays errors when data cannot be refreshed. Project-specific MCP events are read from the project's .atelier/events/ directory. Use the same project path in your coding assistant and monitor to see its activity. Press / to search and ? for the workflow guide. Motion respects the system's reduced-motion preference.

For the optional MCP screenshot and browser audit tools, install Chromium once:

npx playwright install chromium

License

This project is dual-licensed. Pick whichever works for you:

Third-party dependency licenses are fully documented in THIRD_PARTY_LICENSES.md.

Related MCP Connectors

Related MCP Servers

  • F
    license
    A
    quality
    D
    maintenance
    Provides AI assistants with frontend development tools including component scaffolding, bundle analysis, accessibility checks, and responsive design guidance. Enables automated generation of React components with tests and stories, bundle optimization recommendations, and WCAG compliance fixes.
    6
    -
  • F
    license
    A
    quality
    C
    maintenance
    Enables AI assistants to generate production-ready, professional UI design systems and components from simple descriptions, with real images, animated components, and automated quality checks.
    16
    -
  • A
    license
    Not graded
    quality
    B
    maintenance
    Provides AI-powered UI design, repo intelligence, and frontend code generation tools, enabling assistants to scaffold projects, generate and modify UI components, manage design tokens, and analyze repositories.
    MIT
  • F
    license
    Not graded
    quality
    B
    maintenance
    Enables AI coding agents to inspect a project's design tokens and component catalog, audit React/Tailwind files against seven static UI quality rules, score overall UI health, and apply auto-fixes that are re-audited and reverted on regression. It works both locally over filesystem access and remotely on raw TSX/JSX or CSS snippets passed directly as strings.
    -