Draftlytic MCP server

draftlytic-mcp

An MCP server that turns a rough project idea into a structured spec — right inside Claude Code, Cursor, or any MCP-compatible editor. No API key, no account, no network calls. It runs entirely on your machine and hands your editor's model a schema to write into, a checklist of what to ask about, a validator that catches gaps before you start coding, and a renderer that turns the result into a clean Markdown PRD.

This exists because "vibe coding" from a one-line prompt tends to produce a plausible-looking app that's missing half the decisions you actually needed to make — what's in scope for v1, what the data model looks like, what "done" means for a feature. draftlytic-mcp doesn't generate any of that for you; it structures the conversation so your model asks the right questions, then checks its own homework before you start building.

Install

Claude Code

claude mcp add draftlytic -- npx -y draftlytic-mcp

Cursor

Add to .cursor/mcp.json in your project (or the global ~/.cursor/mcp.json):

{
  "mcpServers": {
    "draftlytic": {
      "command": "npx",
      "args": ["-y", "draftlytic-mcp"]
    }
  }
}

Any other MCP client

Most MCP hosts read a generic mcp.json with the same shape:

{
  "mcpServers": {
    "draftlytic": {
      "command": "npx",
      "args": ["-y", "draftlytic-mcp"]
    }
  }
}

Usage

Once connected, ask your editor's model something like:

Use the plan_project prompt for "a habit tracker that reminds me by text message"

It'll walk through specchecklist with you (platform, tech stack, audience, features, competitors, revenue, constraints, data model, notifications, external services, design & UX — a handful of concrete questions per category, many offered as click-to-pick single/multi-select choices rather than blank prompts, with a free-text escape always available), draft a spec, run it through validatespec, fix what comes back, and hand you a rendered PRD in Markdown you can drop straight into a coding-agent prompt, a SPEC.md, or a GitHub issue.

You can also call the tools directly if you already have a spec drafted (by hand, or from another source) and just want it checked and rendered.

Tool reference

Shortened. The full README is on GitHub.