vibe-prompt-mcp

Every vague prompt costs you 2–3 follow-up messages. This MCP fixes your prompt before it reaches the AI — so you get the right output on the first try.

vibe-prompt-mcp scores your prompt across 4 quality dimensions, rewrites the weak parts, and fills in what's missing. The AI gets a precise instruction. You get fewer iterations.

No API key. No account. No server to run. Works inside Claude Code, Cursor, Windsurf, Zed, and any stdio MCP client.

The problem it solves

You send a prompt. The AI produces something close but not quite right. You clarify. It tries again. You say "also add loading states." Another round. "Make it responsive." One more.

Three iterations to get what you could have specified upfront.

vibe-prompt-mcp catches the gaps before the prompt is sent — vague verbs, subjective language, missing acceptance criteria, absent style stack — and fixes them automatically. The AI gets one clear instruction instead of a guessing game.

Related MCP server: PromptArchitect MCP

See it in action

Example 1 — vague UI prompt

Before:

can you please improve the login page, it looks bad and i want it to feel more modern

Score: 76/100 — 5 issues detected

After optimize_prompt:

please redesign the login page, it looks bad and i want it to use
Inter font, neutral color palette, 8px border radius, consistent
16px grid spacing. Done when: the page renders correctly on mobile
and desktop with no console errors. Use Tailwind CSS and shadcn/ui.

Score: 78/100 — filler stripped, vague terms replaced, missing specs appended

Example 2 — feature request with missing specs

Before:

Add a notifications bell icon to the navbar that shows unread count
and a dropdown list of recent notifications with mark-as-read functionality

Score: 79/100 — 3 issues detected

After optimize_prompt:

Add a notifications bell icon to the navbar that shows unread count
and a dropdown list of recent notifications with mark-as-read
functionality. Done when: the list renders correctly on mobile and
desktop with no console errors. Use Tailwind CSS and shadcn/ui.
Include loading, error, and empty states.

Score: 84/100 — acceptance criteria, style stack, and state requirements added

Without this, you'd have built the feature — then asked about loading states, then responsive layout, then the empty state. Three follow-ups eliminated upfront.

How it works

vibe-prompt-mcp runs entirely on your machine as a local Node.js process. It uses no AI, makes no API calls, and sends nothing to any external service.

Under the hood it's a rule engine — 18 rules across 4 dimensions — that analyzes the structure and language of your prompt, detects patterns that consistently cause poor AI output, and applies targeted fixes. Think of it as a linter for prompts.

What this means for you:

• Zero AI cost for the optimization itself. The only tokens spent are ~130 for the tool call overhead (your message + Claude routing the request + the response).

• Net savings come from avoiding re-iterations. Each back-and-forth cycle with the AI costs 500–1,000+ tokens. One optimized prompt that gets it right on the first try pays for itself immediately.

• Runs offline. No network call is made during optimization.

This is the core difference from AI-based prompt improvers — those spend tokens to save tokens. This one doesn't.

Quick start

Step 1 — Add to your AI tool (pick your platform below)

Step 2 — Use it

Ask your AI in plain language:

optimize this prompt: [your prompt here]

score this prompt: [your prompt here]

optimize in verbose mode: [your prompt here]

Step 3 — Send the result

Copy the rewritten prompt and use it as your actual instruction.

Add to your AI tool

Claude Code

Option A — project-level (recommended, checked into source control and shared with your team):

Create .mcp.json at your project root:

{
  "mcpServers": {
    "vibe-prompt-mcp": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "vibe-prompt-mcp"]
    }
  }
}

Option B — global (available in every project on your machine):

claude mcp add vibe-prompt-mcp -s user -- npx -y vibe-prompt-mcp

Restart Claude Code, then type /mcp to confirm vibe-prompt-mcp appears with both tools listed.

Cursor

Settings → MCP → Add new server:

• Name: vibe-prompt-mcp

• Command: npx

• Args: -y vibe-prompt-mcp

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

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

Zed

Add to .zed/settings.json:

{
  "context_servers": {
    "vibe-prompt-mcp": {
      "command": {
        "path": "npx",
        "args": ["-y", "vibe-prompt-mcp"]
      }
    }
  }
}

Antigravity

Add to ~/.gemini/antigravity/mcp_config.json:

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

Lovable / Replit / Codex (HTTP)

These platforms require a deployed remote endpoint. The package ships an HTTP server:

node node_modules/vibe-prompt-mcp/dist/http.js
# Express on port 3000 (or $PORT) — MCP endpoint: POST /mcp

Deploy to Railway or Render and point the platform's MCP URL to https://YOUR_HOST/mcp.

Scoring dimensions

Each dimension is worth 25 points. Total score: 0–100.

Severity:

• 🔴 Critical — high likelihood of wrong output

• ⚠ Warn — reduces quality or causes follow-up iterations

• ✦ Info — noise with no instructional value

Running locally from source

git clone https://github.com/saurabhjambure-pixel/vibe-prompt-mcp
cd vibe-prompt-mcp
npm install
npm run build        # TypeScript → dist/
npm run dev          # stdio server, hot reload
npm run start:http   # HTTP server on port 3000

To use your local build instead of npx:

{
  "mcpServers": {
    "vibe-prompt-mcp": {
      "command": "node",
      "args": ["/path/to/vibe-prompt-mcp/dist/index.js"]
    }
  }
}

Contributing

Issues and PRs welcome. If you have a rule idea — a pattern you keep seeing that produces poor AI output — open an issue.

License

MIT