Prompt Suggestion MCP Server (Claude Desktop – Local Version)

A Model Context Protocol (MCP) server that helps your LLM instantly rewrite disappointing responses into one clean, high-quality prompt — running fully on your local machine.

This MCP server analyzes the recent conversation, identifies what you actually wanted, and generates a single, first-person improved prompt designed to produce a much better answer from the assistant.

No noise.
No confusion.
Just one perfect next prompt — every time.

---

✨ Features

🎯 Smart Failure Understanding

The server reads:

• The recent conversation window
• The last user message
• An optional user comment explaining the dislike

It then determines:

• What you were trying to achieve
• Why the assistant’s last answer failed
• What clarity or structure was missing

---

✍️ High-Quality Prompt Rewriting

Your MCP server always generates one optimized prompt that is:

• Written in first person
  • (“Explain to me…”, “Give me…”, “Help me…”)
• Self-contained
  • Works even if the AI never saw the earlier conversation
• Clear, specific, and free of meta references
  • No “the user”, “previous messages”, or “the conversation above”
• Structured when needed
  • Steps, examples, analogies, UX guidelines, etc.

You may also receive up to two short alternatives, but the main one is always clearly the best.

---

🧠 Consistent Output

No matter what phrase you use:

• “use the tool”
• “give me a better prompt”
• “rewrite this prompt”
• “fix the prompt”

…the MCP server always returns one definitive best prompt.

---

🖥️ Local Claude Desktop Version (This Repo)

This repository contains the local MCP server (TypeScript version) used with Claude Desktop.

It runs fully on your machine using tsx — no cloud hosting required.

---

📦 Installation

1️⃣ Clone the repository

git clone https://github.com/d180/PromptSuggestion-Claude-Desktop.git
cd PromptSuggestion-Claude-Desktop

2️⃣ Install dependencies

npm install

3️⃣ Add your API Key

Create a .env file inside the project root:

GEMINI_API_KEY=your_key_here

That’s all you need locally.

---

🔗 Connect the MCP Server to Claude Desktop

Claude Desktop loads MCP connections from this file:

macOS

~/Library/Application Support/Claude/claude_desktop_config.json

Windows

%APPDATA%/Claude/claude_desktop_config.json.json

Add this configuration:

{
  "mcpServers": {
    "prompt-suggestion": {
      "command": "npx",
      "args": [
        "tsx",
        "PATH_TO_YOUR_PROJECT/src/stdio.ts"
      ],
      "env": {
        "GEMINI_API_KEY": "YOUR_API_KEY_HERE"
      }
    }
  }
}

Replace:

• PATH_TO_YOUR_PROJECT → Your actual local folder path
• YOUR_API_KEY_HERE → Your Gemini API key

Then restart Claude Desktop.

Your tool will appear automatically when Claude asks for MCP permissions.

---

🧠 How to Use the MCP Tool

When chatting with Claude, simply type any of the following:

• "Use the tool and give me a better prompt."
• "Rewrite this prompt using the MCP tool."
• "Fix this prompt."
• "Give me a cleaner prompt for this."

Claude Desktop will:

1. Detect the MCP intent
2. Call your local MCP server
3. Request permission to run it
4. Return the improved prompt

---

🧩 Output Format

The MCP server always returns:

{
  "summary": "...",
  "root_causes": ["..."],
  "suggested_prompt": "THE BEST FIRST-PERSON PROMPT",
  "alternatives": ["..."],
  "confidence": "0.0 - 1.0"
}

The best improved prompt is always in:
"suggested_prompt"

---

🚀 Why the Local Version?

• No ChatGPT developer mode required
• Claude’s memory stays ON
• Runs entirely on your machine
• No cloud, no accounts, no hosting
• Extremely fast and private

---