cli2skill (2233admin)

🔧 Turn any MCP into an Agent Skill — zero MCP overhead

把任意 MCP 变成 Agent Skill —— 告别 MCP 进程泄漏

Quick Start · Why Not MCP · Examples · Full Story (血案全文)

The Problem · 问题

EN: 15 MCP servers accumulated 400+ zombie processes on a 32GB Windows machine, exhausting virtual memory commit charge (0xC000012D). Every fork() failed. The system was completely dead.

CN: 15 个 MCP server 在 32GB Windows 机器上积累了 400+ 僵尸进程，耗尽虚拟内存 commit charge（0xC000012D），所有 fork() 失败，系统彻底死机。

Each MCP server spawns a 3-layer process tree that never exits:

每个 MCP server 产生一个三层进程树，session 结束后不回收：

cmd.exe          ← shell wrapper / 壳
  └── uv.exe       ← Python launcher / 启动器
       └── python.exe  ← MCP server / 服务进程

× 15 servers = 45+ persistent processes → system dead in 5-10 sessions

*"Software as a tool 结论是别用 MCP... 这玩意是拿去忽悠子刊审稿人用的"* *—— 某大厂 RL 团队工程师的非公开评论*

Bug report: anthropics/claude-code#38228

The Solution · 解决方案

Don't run a server. Just run a command.

别跑服务器，跑命令就行。

# Before (MCP): persistent process, leaks memory, hard to debug
# 之前 (MCP): 常驻进程，泄漏内存，难调试
Agent --JSON-RPC--> MCP Server (persistent) --API--> Backend

# After (CLI + Skill): runs, returns, exits
# 之后 (CLI + Skill): 跑完退出，零开销
Agent --Bash tool--> CLI (on-demand) --API--> Backend

cli2skill parses any CLI's --help and generates an Agent Skill — a markdown file that teaches your AI agent when and how to call your tool.

cli2skill 解析任意 CLI 的 --help，生成 Agent Skill —— 一个教 AI agent 何时、如何调用你工具的 markdown 文件。

Works with 26+ platforms · 兼容 26+ 平台： Claude Code · Codex CLI · Gemini CLI · Cursor · Copilot · VS Code...

Quick Start · 快速开始

# ClawHub (recommended for Claude Code users)
clawhub install cli2skill

# PyPI
pip install cli2skill

# Generate a skill from any CLI / 从任意 CLI 生成 Skill
cli2skill generate gh --name github-cli -o ~/.claude/skills/

# Preview parsed metadata / 预览解析结果
cli2skill preview mytool

# Custom executable path / 自定义可执行路径
cli2skill generate "python my_tool.py" --name my-tool \
  --exe-path "python /full/path/my_tool.py"

# Parse from saved help text / 从文件解析
cli2skill generate mytool --help-file help_output.txt

That's it. Drop the generated .md into ~/.claude/skills/ (or your platform's skill directory), and your agent can use the CLI immediately.

就这样。把生成的 .md 放进 ~/.claude/skills/（或你平台的 skill 目录），agent 立刻就能用。

Why Not MCP · 为什么不用 MCP

*\* Anthropic's own benchmark: code execution vs direct MCP tool calls = 98.7% token savings.*

When MCP IS right · MCP 仍然适用的场景

• 🟢 Persistent browser sessions / 持久化浏览器会话
• 🟢 Multi-client shared servers / 多客户端共享服务
• 🟢 Streaming notifications / 流式通知推送
• 🟢 Remote HTTP endpoints (Slack, GitHub, Notion) / 远程 HTTP 端点

Rule of thumb / 经验法则: If your tool is "call → return result", use CLI + Skill. If it needs persistent state or push notifications, MCP is fine.

如果你的工具是"调用 → 返回结果"，用 CLI + Skill。如果需要持久状态或推送通知，MCP 是对的。

Examples · 示例

Generated: GitHub CLI

cli2skill generate gh --name github-cli --no-subcommands

Output (click to expand)

---
name: github-cli
description: Work seamlessly with GitHub from the command line.
user-invocable: false
allowed-tools: Bash(gh *)
---

# github-cli

## Commands

gh auth
gh browse
gh issue
gh pr
gh repo
gh release
...

## When to use
- Work seamlessly with GitHub from the command line.
- Available commands: `auth`, `browse`, `issue`, `pr`, `repo`

Generated: Custom Python Tool

cli2skill generate "python memu-cli.py" --name memu \
  --exe-path "python /path/to/memu-cli.py"

Output (click to expand)

---
name: memu
description: memU memory CLI
user-invocable: false
allowed-tools: Bash(python /path/to/memu-cli.py *)
---

# memu

## Commands

python /path/to/memu-cli.py search
python /path/to/memu-cli.py query
python /path/to/memu-cli.py add
python /path/to/memu-cli.py categories

### search
Semantic search

### query
Search + categories

### add
Memorize content

### categories
List categories

Real Migration: 4 MCP Servers → 0 Persistent Processes

实际迁移：4 个 MCP Server → 0 个常驻进程

Supported Formats · 支持的格式

How It Works · 工作原理

graph LR
    A["Your CLI
你的 CLI"] -->|--help| B["cli2skill
解析器"]
    B -->|parse| C["CLIMetadata
命令元数据"]
    C -->|generate| D["SKILL.md
Agent Skill"]
    D -->|drop in| E["~/.claude/skills/
技能目录"]
    E -->|auto-discover| F["Agent uses CLI
Agent 调用 CLI"]

Background · 背景

This tool was born from a real production incident on 2026-03-24. Full technical writeup with process snapshots, root cause analysis, and references:

这个工具诞生于 2026-03-24 的一次真实生产事故。完整技术分析（含进程快照、根因分析、引用）：

📖 Read the full story → ARTICLE.md

References · 参考

• Agent Skills Specification — Anthropic, 2025-12
• Code Execution with MCP — Anthropic (98.7% token savings)
• Skills vs Dynamic MCP Loadouts — Armin Ronacher
• CLI-Anything — HKU, 22.3k stars
• anthropics/claude-code#38228 — MCP process leak bug

License

MIT · Zero dependencies · Python 3.10+