@neuraldraft/mcp directly. Same conventions, less
copy-paste.
Copy this
Why this works
The model’s job becomes mechanical: every text node gets adata-translate
attribute, every image a data-image-key, every fetch routes to a real URL.
On the Neural Draft side, the first time a key is seen we auto-create it as
empty; the customer fills it in via the admin. No schema work, no migrations,
no glue.
If your tool also supports a “knowledge” or “files” section, paste the OpenAPI
URL — https://api.neuraldraft.io/v1/openapi.yaml — and the model will
ground its API calls in the spec.
Hosted-tool checklist
For the three native MCP clients (Cursor / Claude Code / Continue), use the
MCP setup guide instead — it gives the AI tool access to the
real tools (
register_component, generate_blog_post, generate_image,
etc.), not just rules.
Working example — Lovable system prompt + SDK glue
If your Lovable project has a “Server functions” / Node runtime, drop this snippet into a server file. Once Lovable adds MCP, the same code keeps working.“Build a marketing home page using the Neural Draft conventions in my system prompt. Hero, three-feature grid, CTA, footer. Every text node needsdata-translate="<key>". Every image needsdata-image-key="<key>". UseloadHomepageMeta()fromlib/neural-draft.tsfor<title>and meta tags. Default copy in the brand voice listed in my system prompt.”