Meta Prompt Mcp

io.github.kapillamba4由 GitHub 账号 kapillamba4 发布v1.0.11更新于 Oct 24, 2026

MCP server providing official Google and Anthropic prompting guides for meta-prompt generation.

已验证STDIO仅桌面知识与记忆AI 与机器学习

概览

AI 生成的概览

一个本地 MCP 服务器,提供内置的 Google 与 Anthropic 提示词指南,让助手依据官方最佳实践撰写元提示词和系统提示词。

可以做什么
从内置的 markdown 文件中读取 Google 提示词指南(涵盖技巧、最佳实践与 LLM 配置)以及 Anthropic 关于思维链、多示例提示和扩展思考的指南,全程无网络调用。助手阅读这些指南后,可据此为特定任务提出改进建议或选择合适技巧。
适用场景
适合在编写或优化智能体的系统提示词与元提示词时使用,或当你希望提示词建议基于 Google 与 Anthropic 的方法论而非凭空猜测时使用。
运行要求
通过 stdio 在本地运行;可用 uvx 或 pip 安装 Python 包 meta-prompt-mcp。日常使用无需 API 密钥或网络访问。可选的基准测试目标需要 OPENROUTER_API_KEY。
安装前请注意
该服务器只读取内置的静态指南,不写入或发送数据。仅可选的基准测试步骤会调用外部服务并需要 OPENROUTER_API_KEY,应避免将其放入共享配置中。

安装

在 SourceWeft 中

  1. 打开 控制台中的 Meta Prompt Mcp,将其添加到工作区。
  2. 为需要使用其工具的对话启用该服务。

Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。

其他 MCP 客户端

参照 仓库 中的启动说明。

README

Meta-Prompt MCP

Instant access to Google and Anthropic's official prompting guides within your LLM workflow—optimized for crafting high-quality meta-prompts and system prompts.

[Python 3.10+] [License: MIT]


What It Does

Meta-Prompt MCP is an MCP server that surfaces official prompting best practices from Google and Anthropic directly within your LLM workflow. Instead of searching documentation or guessing how to instruct an LLM, you can query expert guides on-demand.

This is especially valuable when crafting meta-prompts (system prompts for agents) or optimizing your own prompting strategies. By grounding your prompts in proven methodology, you'll generate more effective, well-structured outputs.

How It Works

Your MCP host (Claude Desktop, Cursor, etc.) communicates with the server via stdio. When you ask for a guide, the server retrieves it from bundled markdown files—no API calls, no latency.

┌────────────────────────┐          ┌──────────────────────────┐│   Your LLM Host        │◄─stdio──►│  Meta-Prompt MCP Server  ││  (Claude Desktop, etc) │          │                          │└────────────────────────┘          │  • get_google_guide      │                                    │  • get_anthropic_guide   │                                    │                          │                                    │  Data layer:             │                                    │  ./data/                 │                                    │  ├── google_*.md         │                                    │  └── anthropic_*.md      │                                    └──────────────────────────┘

Key Features

  • get_google_guide — Retrieves Google's comprehensive prompting guide covering techniques, best practices, and LLM configuration
  • get_anthropic_guide — Retrieves Anthropic's guide on chain-of-thought, multishot prompting, and extended thinking
  • Zero dependencies — Runs entirely offline with bundled guides; no API keys or network calls required

Validation

We ran a benchmark comparing prompts generated with and without guide access across 5 diverse tasks. An independent judge LLM scored each on Clarity, Specificity, Structure, Effectiveness, and Overall quality (1–10 scale).

See full results →

To reproduce the benchmark:

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

Quick Start

1. Install

bash
# Via uvx (recommended — run without installing globally)uvx meta-prompt-mcp
# Or install via pippip 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:

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

Add to your MCP settings:

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

Run the following command in your terminal:

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

Usage

Query the guides while crafting your prompts. Examples:

  • "I'm building a code reviewer agent. Reference the Google guide to help me write a better system prompt."
  • "Based on the Anthropic guide, suggest improvements to this prompt for better reasoning."
  • "What technique from the guides would work best for this task?"

The LLM reads the guides and uses that knowledge to give you more informed suggestions and feedback on your prompts.


Development

bash
# Clone the repogit clone https://github.com/kapillamba4/meta-prompt-mcp.gitcd meta-prompt-mcp
# Install in dev modemake dev
# Run the servermake run

Available Commands

bash
make dev        # Install in editable mode with dev dependenciesmake run        # Start the MCP server locallymake lint       # Check code qualitymake format     # Auto-format code with Black & isortmake test       # Run test suitemake benchmark  # Run prompt quality benchmark (requires OPENROUTER_API_KEY)make build      # Build distribution packagesmake publish    # Publish to PyPI

Project Structure

meta-prompt-mcp/├── pyproject.toml                 # Dependencies and package config├── Makefile                       # Development commands├── .env.example                   # Template for OPENROUTER_API_KEY│├── benchmarks/│   ├── benchmark.py               # Prompt quality evaluation script│   └── results.md                 # Benchmark results│└── src/meta_prompt_mcp/    ├── __init__.py    ├── __main__.py                # Entry point (python -m)    ├── server.py                  # FastMCP server & tool definitions    └── data/        ├── google_prompting_guide.md        └── anthropic_prompting_guide.md

Contributing

Issues, feature requests, and contributions welcome. Please open an issue on GitHub.

License

MIT

来源:README.md,提交 3b8d06a

工具

0
SourceWeft 尚未在沙箱中启动过该软件包,因此其工具未知。

版本历史

1
  1. v1.0.11最新Sep 16, 2026

评分

已满足适用检查项的 50%

该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。

  • 近 30 天内可连通25 分

    请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。

  • 已声明认证方式10 分

    请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。

  • 已列出工具15 分

    请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。

  • 描述必需15 分
  • 图标10 分
  • 文档15 分
  • 许可证10 分