Meta-Prompt MCP

A Prompting Oracle — An MCP server that bridges official Prompting Guides with your LLM workflow to help you generate highly accurate, effective, and structured meta-prompts.

---

What It Does

Meta-Prompt MCP is a specialized Model Context Protocol (MCP) server that acts as an automated "Prompting Oracle." It empowers any MCP-compatible host (Claude Desktop, Cursor, etc.) to query expert Prompting Guides mid-conversation.

When building AI workflows, creating robust "meta-prompts" (system prompts for agents) is critical. Instead of guessing how to instruct an LLM, this server provides immediate access to authoritative guidelines. By surfacing these best practices on-demand, it ensures the meta-prompts you generate are exceptionally accurate, helpful, and grounded in proven methodology.

Architecture

┌─────────────────────┐     stdio      ┌──────────────────────────┐
│   MCP Host          │◄──────────────►│   Meta-Prompt MCP        │
│   (Claude Desktop,  │                │                          │
│    Cursor, IDEs)    │                │   ┌──────────────────┐   │
│                     │                │   │  FastMCP Server   │   │
│                     │                │   │  • get_google_    │   │
│                     │                │   │    guide          │   │
│                     │                │   │  • get_anthropic_ │   │
│                     │                │   │    guide          │   │
│                     │                │   └────────┬─────────┘   │
│                     │                │            │              │
│                     │                │   ┌────────▼─────────┐   │
│                     │                │   │  ./data/         │   │
│                     │                │   │  (markdown files) │   │
│                     │                │   └──────────────────┘   │
└─────────────────────┘                └──────────────────────────┘

Key Features

Feature	Details
get_google_guide tool	Retrieves the comprehensive Google Prompting Guide to inform clear, context-rich prompting strategies
get_anthropic_guide tool	Retrieves the full Anthropic Prompting Guide for mastering capabilities and system prompts
Offline capable	Runs entirely locally, reading from bundled markdown files with zero API dependencies

---

Benchmark Results

To validate the tool's impact, we ran a benchmark comparing prompts generated with and without the prompting guides across 5 diverse tasks. An independent judge LLM scored each prompt on Clarity, Specificity, Structure, Effectiveness, and Overall quality (1–10 scale).

View Full Benchmark Results →

Run the benchmark yourself:

export OPENROUTER_API_KEY=sk-or-...
make benchmark

---

Quick Start

1. Install

# Via uvx (recommended — run without installing globally)
uvx meta-prompt-mcp

# Or install via pip
pip install meta-prompt-mcp

The package ships with bundled markdown guides — no API keys or setup needed.

2. Configure Your MCP Host

Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "meta-prompt-mcp": {
      "command": "uvx",
      "args": ["meta-prompt-mcp"]
    }
  }
}

Cursor

Add to your MCP settings:

{
  "mcpServers": {
    "meta-prompt-mcp": {
      "command": "uvx",
      "args": ["meta-prompt-mcp"]
    }
  }
}

Claude Code

Run the following command in your terminal:

claude mcp add meta-prompt-mcp -- uvx meta-prompt-mcp

---

Development

# Clone the repo
git clone <your-repo-url>
cd meta-prompt-mcp

# Install in dev mode
make dev

# Run the server
make run

Make Targets

Command	Description
make dev	Install in editable mode with dev dependencies
make run	Start the MCP server
make benchmark	Run prompt quality benchmark (requires OPENROUTER_API_KEY)
make lint	Run linter
make format	Auto-format code
make test	Run tests
make build	Build distribution packages
make publish	Publish to PyPI

---

Project Structure

meta-prompt-mcp/
├── pyproject.toml              # Package config & dependencies
├── Makefile                    # Dev commands (make help)
├── README.md
├── .env.example                # Env template (OPENROUTER_API_KEY)
├── benchmarks/
│   ├── benchmark.py            # Prompt quality benchmark
│   └── results.md              # Generated benchmark results
└── src/
    └── meta_prompt_mcp/
        ├── __init__.py
        ├── __main__.py         # python -m support
        ├── server.py           # FastMCP server with tools
        └── data/               # Bundled markdown guides

---

License

MIT