Neural Draft

@neuraldraft/mcp

Make AI-built sites real businesses. Plug Neural Draft's CMS, blog, social, booking, and commerce APIs into Claude Code, Cursor, Continue, and any other MCP client.

What it does

@neuraldraft/mcp is a Model Context Protocol server that gives AI coding tools first-class access to the Neural Draft backend platform. When your AI assistant builds a site, it can:

• read your project's brand context (voice, colors, fonts, audience)
• register every section it generates as an editable component in your admin
• create translation keys, kick off blog/image generation, list products, set up booking widgets
• emit markup that's CMS-managed by default — no manual wiring

Demo

A real Claude Code session driving Neural Draft via MCP. One natural-language prompt creates 4 translation keys (EN + HR) and registers a hero image.

Watch the full MP4 →

Why it exists

AI codegen tools (Lovable, Claude Code, v0, Bolt, Cursor) are excellent at building frontends. They're terrible at the bits that make a frontend a real business — multi-language CMS, blog pipelines, social scheduling, bookings, e-commerce. Neural Draft is the backend that handles those. This MCP server is the wire between the two: the AI tool generates code that uses Neural Draft correctly by default, because the conventions and tools are exposed as MCP context.

The pitch: "Build your site with Lovable. Run the actual business on Neural Draft."

Quick install

You need a Neural Draft project API key. Get one at neuraldraft.io/dashboard/api-keys. Test-mode keys (ndsk_test_…) work out of the box.

Claude Code

~/.config/claude-code/mcp.json:

{
  "mcpServers": {
    "neuraldraft": {
      "command": "npx",
      "args": ["-y", "@neuraldraft/mcp"],
      "env": {
        "NEURALDRAFT_API_KEY": "ndsk_live_xxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Or use the CLI:

claude mcp add neuraldraft \
  -e NEURALDRAFT_API_KEY=ndsk_live_xxx \
  -- npx -y @neuraldraft/mcp

Cursor

.cursor/mcp.json (workspace) or ~/.cursor/mcp.json (global):

{
  "mcpServers": {
    "neuraldraft": {
      "command": "npx",
      "args": ["-y", "@neuraldraft/mcp"],
      "env": {
        "NEURALDRAFT_API_KEY": "ndsk_live_xxxxxxxxxxxxxxxxxxxxxxxx"
      }
    }
  }
}

Continue

~/.continue/config.yaml:

experimental:
  modelContextProtocolServers:
    - transport:
        type: stdio
        command: npx
        args: ["-y", "@neuraldraft/mcp"]
      env:
        NEURALDRAFT_API_KEY: ndsk_live_xxxxxxxxxxxxxxxxxxxxxxxx

Then restart your editor. You should see neuraldraft listed under MCP servers, with resources, tools, and prompts available.

Auth setup

The server reads two environment variables (set by your MCP client config — never commit them):

NEURAL_DRAFT_* (with underscore) is also accepted as an alias for any of the above.

What it exposes

Resources (read-only context)

Tools (functions the AI can call)

Prompts (slash commands)

Examples

1. "Build me a blog page"

You: /scaffold-blog-page framework=next language=en

Claude: Reading brand://current… (voice: warm and grounded; audience: yoga-curious adults)
Claude: Reading schema://blog-post… (got it — id, slug, title, content, featured_image, tags, …)
Claude: Reading conventions://api-usage… (auth via Bearer ndsk_…, rate limit 60/min, RFC 7807 errors)

Claude: Generating /lib/neuraldraft.ts…
Claude: Generating /app/blog/page.tsx (index)…
Claude: Calling register_component { intent: 'blog_index_card', html: ... }
Claude: → cmp_2NgB7hKqL (edit at https://app.neuraldraft.io/c/cmp_2NgB7hKqL)

Claude: Generating /app/blog/[slug]/page.tsx…
Claude: Calling register_component { intent: 'blog_post_meta_block', html: ... }
Claude: → cmp_2NgC9qPLq (edit at https://app.neuraldraft.io/c/cmp_2NgC9qPLq)

Claude: Done. 2 components registered. Open the editor URLs to change copy.

2. "Add a hero with my brand colors"

You: Build a hero section that says "Find your calm" with a CTA to /book.

Claude: Reading brand://current… (deep sage #3F6B5C primary, linen #E8DCC4 secondary, Fraunces / Inter)
Claude: Reading conventions://editable-html… (every text node needs data-translate)

Claude: Calling generate_image { prompt: "Serene yoga studio at dawn, sage and linen tones, no people", aspect_ratio: "16:9", key: "hero.background" }
Claude: → job_img_7K9L (poll with get_job)

Claude: Drafting <section class="hero">…</section> with data-translate="hero.headline" / "hero.subhead" / "hero.cta"
Claude: Calling register_component { intent: "marketing_hero", page_slug: "home", html: ... }
Claude: → cmp_2NgD9KqL — translation keys created: hero.headline, hero.subhead, hero.cta.

Claude: I'll poll the image job…
Claude: Calling get_job { id: "job_img_7K9L" } → status: completed, url: https://cdn.neuraldraft.io/.../hero-bg.jpg

Claude: Done. The hero is in your admin (https://app.neuraldraft.io/c/cmp_2NgD9KqL) — you can change the headline, swap the image, or recolor without touching code.

3. "Set up a storefront for the products I already have"

You: I have 14 products in my Neural Draft project. Scaffold a /shop page.

Claude: Calling list_products { page_size: 20 } → 14 products returned (page 1/1).
Claude: Reading schema://product… (id, name, price_cents, currency, status, images, …)
Claude: Reading brand://current… (warm voice; primary #3F6B5C)
Claude: Reading conventions://editable-html…

Claude: Generating /app/shop/page.tsx with a grid of 14 ProductCard components.
Claude: Calling register_component { intent: "product_grid_header", page_slug: "shop", html: ... } → cmp_2NgE
Claude: Generating /app/shop/[slug]/page.tsx with get_product fetch + buy-button widget.
Claude: Calling register_component { intent: "product_detail_layout", page_slug: "shop_detail", html: ... } → cmp_2NgF

Claude: Done. The grid auto-renders your 14 products (real names, prices, images). Edit the page header in admin: https://app.neuraldraft.io/c/cmp_2NgE.

Local development

git clone https://github.com/vbalagovic/neuraldraft-mcp.git
cd mcp
npm install
npm run lint    # tsc --noEmit
npm test
npm run build   # → dist/

Run against a local API:

NEURALDRAFT_API_KEY=ndsk_test_xxx \
NEURALDRAFT_API_URL=http://localhost:8080/v1 \
npm run dev

Inspect interactively with the official MCP Inspector:

npx @modelcontextprotocol/inspector node dist/index.js

(Or npx @modelcontextprotocol/inspector --cli node dist/index.js --tool register_component --args '{"html":"<h1>Hi</h1>","intent":"hero"}' for one-shot CLI calls.)

Architecture

This server is a thin wrapper over the Neural Draft Project API v1. Every tool maps to one (or a handful of) REST calls; resources are either constants (schemas, conventions docs) or cached reads (brand://current). The full API spec lives at openapi.yaml in the platform repo.

Stdio is the only transport in v0.x — every current AI coding tool spawns local processes. Streamable HTTP transport will arrive when there's hosted multi-tenant demand.

Contributing

Issues and PRs welcome. See open issues for places to start. PRs should:

• Keep tests at ≥ 80% line coverage on src/
• Use the existing register* helper pattern when adding tools/resources
• Never write to stdout in server runtime code (stdio mode breaks)
• Update the README's "What it exposes" table for any new surface

Changelog

See CHANGELOG.md (TBD — this is the first public release).

License

MIT — open-source from day one.

Built by Neural Draft. Questions? developers@neuraldraft.io.