MCP overview

The Model Context Protocol (MCP) is the missing wire between an AI coding tool and a real backend. Without it, your AI knows what JavaScript looks like but hallucinates your endpoints. With it, your AI knows your project’s brand context, your data shapes, your conventions — and uses real tools to make real changes. @neuraldraft/mcp is our open-source MCP server. Drop it into Claude Code, Cursor, or Continue and your AI starts generating Neural Draft-native code automatically. No reading the docs. No glue code.

​

Why this matters

Every AI codegen tool today (Lovable, Claude Code, Cursor, v0, Bolt) is great at generating frontends and bad at integrating with anything that isn’t on its training data. The output is usually a beautiful page that calls /api/some-endpoint-that-doesnt-exist. MCP fixes that with three primitives. Resources, tools, and prompts — each designed to give the AI just enough live context to be correct without overwhelming the prompt:

• Resources are read-only context. The AI loads brand://current once per session and now knows your colors, fonts, voice, and audience.
• Tools are functions the AI can call. When it generates a hero section, it calls register_component(html, intent) — and the section appears in your admin as editable.
• Prompts are slash commands the user can invoke. /scaffold-marketing-site drops a complete, opinionated, Neural-Draft-native scaffold into the current project.

The result is “compliance by default” — every section gets data-translate attributes; every page reads brand colors automatically; every API call hits the real Neural Draft endpoint instead of a hallucinated one.

​

What @neuraldraft/mcp exposes

​

Resources

​

Tools

​

Prompts (slash commands)

​

How the magic works

When a user types “build a marketing hero” in Claude Code with Neural Draft’s MCP active, three things happen invisibly:

1. Context injection — the AI reads brand://current and now uses your actual colors and voice. It reads conventions://editable-html and now knows every text element should be data-translate="...".
2. Tool-driven registration — the AI generates the hero, then calls register_component(html, intent). The component appears in your admin immediately, with the data-translate keys auto-discovered as editable fields.
3. Brand auto-application — the AI doesn’t need you to say “use our purple” or “match our voice” again. The brand resource is in context for every subsequent section.

You never had to ask the AI to do any of that. Most users never realise it happens.

​

Open source from day one

@neuraldraft/mcp is Apache-2.0 licensed and published on npm. The package is a thin wrapper around the v1 Project API — same key, same endpoints, same behaviour as if you’d called REST directly. The companion TypeScript SDK is on GitHub at vbalagovic/neuraldraft-sdk; file MCP bugs there until the standalone repo lands. We maintain both because they’re the most important onboarding assets in the platform.

​

What about Lovable, v0, Bolt?

These hosted codegen tools don’t yet support MCP — they’re closed-loop. For those environments, we ship a system-prompt block (frameworks/lovable) you paste into the build prompt. It teaches the model the same conventions the MCP server enforces. When Lovable/Bolt eventually ship MCP support, we’re already there — the server we ship today will work without changes.

​

Next steps