@aicanvas/mcp

Model Context Protocol server for AI Canvas. Lets your AI editor (Claude Code, Cursor, Claude Desktop, Codex) discover and install AI Canvas components without leaving the chat.

You:   "find me an animated card stack"
AI:    [calls search_components → ranks matches → picks polaroid-stack]
       [calls get_install_command → returns: npx shadcn add @aicanvas/polaroid-stack]
       [runs the command — file lands in your project]
You:   "make the rotation more dramatic"
AI:    [edits the file directly]

What it exposes

Thirteen tools, derived from the live aicanvas.me registry. Adding a component to the website auto-updates the MCP within minutes. The registry holds 80+ standalone components, plus the Andromeda design system with its own components and templates.

Tool	Purpose
list_categories	Show every category with component counts. Orient before drilling in.
list_components	Browse standalone components, optionally filtered by category. Paginated.
search_components	Fuzzy keyword search across standalone AND design-system components (slugs, names, descriptions, categories, tags). Ranked results.
get_component	Full metadata plus complete .tsx source code for one component. Resolves both standalone slugs and design-system component slugs (e.g. andromeda-heat-grid).
get_install_command	Get the npx shadcn add @aicanvas/<slug> command for any standalone or design-system component.
list_systems	List the design systems available (e.g. Andromeda), with component and template counts.
get_system	Every file for a whole design system, plus its shared tokens, its component install commands, and the dependency chain.
list_templates	List the ready-made templates (complete example screens), optionally filtered by design system.
get_template	Every file for a single template (a ready-made composition), with the registry dependencies it pulls in.
get_component_props	The documented props of one component: name, type, required, default, description. Free, no account, far smaller than the source.
validate_usage	Check a file's JSX against the documented props of the AI Canvas components it imports: unknown props, missing required props, values outside a prop's options, each with a line number.
compose_page	Turn a one-line brief into a build plan: closest template, one component per part with alternatives, install order (tokens first), and the parts nothing covers. Plans only, generates no code.
get_audit_checklist	The checklist to run on a finished page or component: install completeness, props, both themes, 320px layout, motion cleanup, keyboard and contrast, composition.

All tools are read-only. Nothing is mutated on AI Canvas's side.

The workflow the last four tools are built for: compose_page to plan, get_component_props to read each API before writing JSX, validate_usage on the finished file, get_audit_checklist before calling it done.

Standalone components and design-system components are kept in separate buckets. list_components and list_categories cover the 80+ standalones only, so the catalog stays clean. Design-system components are reached through list_systems / get_system, and are also resolvable by slug through get_component, get_install_command, and search_components.

Setup

Claude Desktop

Edit claude_desktop_config.json:

• macOS: ~/Library/Application Support/Claude/claude_desktop_config.json
• Windows: %APPDATA%\Claude\claude_desktop_config.json
• Linux: ~/.config/Claude/claude_desktop_config.json

{
  "mcpServers": {
    "aicanvas": {
      "command": "npx",
      "args": ["-y", "@aicanvas/mcp"]
    }
  }
}

Restart Claude Desktop. The aicanvas tools should appear in the tools list.

Cursor

Project-level (.cursor/mcp.json in your repo) or global (~/.cursor/mcp.json):

{
  "mcpServers": {
    "aicanvas": {
      "command": "npx",
      "args": ["-y", "@aicanvas/mcp"]
    }
  }
}

Claude Code

Project-level .mcp.json (committed alongside your code):

{
  "mcpServers": {
    "aicanvas": {
      "command": "npx",
      "args": ["-y", "@aicanvas/mcp"]
    }
  }
}

Or via the CLI:

claude mcp add aicanvas -- npx -y @aicanvas/mcp

Codex

~/.codex/config.toml:

[mcp_servers.aicanvas]
command = "npx"
args = ["-y", "@aicanvas/mcp"]

Verify it works

Inside any of the AI editors above, ask:

"What categories of components does AI Canvas have?"

You should see all 11 categories with counts. If you don't, check the host's MCP logs. Typically a "Failed to spawn process" or network error means npx couldn't fetch the package, or the host wasn't restarted after editing the config.

To inspect the server outside an editor:

npx -y @modelcontextprotocol/inspector npx -y @aicanvas/mcp

This opens a browser UI where you can call each tool by hand.

Configuration

Env var	Default	What it does
AICANVAS_REGISTRY_BASE	https://aicanvas.me/r	Registry root URL. Override only for local development against a self-hosted mirror.
AICANVAS_TOKEN	(none)	Optional per-account token. The website bakes it into the MCP config it gives you. It identifies your AI Canvas account so source pulls are authorized and any premium content you own unlocks. Sent as a Bearer token on every registry request. Without it you are anonymous: metadata browsing still works, but pulling a component's source returns a short "create a free account" notice instead of the code.

Accounts and premium content

Browsing metadata (list_categories, list_components, search_components, list_systems, list_templates, get_install_command) is free and needs no account. Pulling actual source (get_component, get_system, get_template) runs against a free AI Canvas account: set AICANVAS_TOKEN from your account and the website gives you the ready-made config. Without a token, a free-component source pull returns a short "create a free account" notice instead of the code, and premium content (full design systems and templates) returns HTTP 402 with an upgrade link at aicanvas.me/pricing. Update anytime with npx @aicanvas/mcp@latest.

How the registry stays fresh

The MCP fetches https://aicanvas.me/r/aicanvas-mcp.json on first use and caches it for 5 minutes. New components on AI Canvas appear in the MCP automatically, no package update required.

The full source code for each component is fetched on demand from /r/<slug>.json (the same shadcn registry items the CLI installs).

The prop tables come from https://aicanvas.me/r/aicanvas-props.json, parsed at build time from each component's own JSDoc, so they cannot drift from the source. Same 5-minute cache.

Local development

git clone https://github.com/aicanvas-me/aicanvas.git
cd aicanvas/mcp
npm install
npm run build

# Smoke-test against the deployed registry:
npm run inspect

# Or point at a locally running aicanvas dev server:
AICANVAS_REGISTRY_BASE=http://localhost:3000/r node dist/index.js

Edit src/index.ts, run npm run build, restart your AI editor.

License

MIT, same as the AI Canvas project.