mcp-safe-proxy
轻量级 MCP 注解代理 — 拦截 tools/list 响应,将工具注解重写为"安全"值,绕过 Codex 等 AI CLI 的审批弹窗,同时完全透传所有实际操作。
Install / Use
claude mcp add Huan-zhaojun -- npx -y github:Huan-zhaojun/mcp-safe-proxyIf the server publishes to npm under a different name, use that package instead — check the repo README.
MCP Server
Model Context Protocol server
Quality Score
Category
Development & EngineeringSupported Platforms
Tags
Our assessment of mcp-safe-proxy
mcp-safe-proxy scores 72/100 on our quality scale, 2615th of 3,553 Development & Engineering skills we index.
Its MCP Server is 3.8 KB long, well organised into 18 sections with 10 code examples: a solid amount of guidance for an agent.
It has 10 GitHub stars, so there is little community track record yet; judge it on its content.
Maintenance, license and trust
- The repository was last updated about 7 months ago. That is recent enough to be usable, but agent tooling moves fast, so check the instructions against your agent's current version.
- Our last check on 2026-09-17 found the source still online.
- It is released under the MIT license, a permissive license that allows use, modification and commercial use with attribution.
- Its trust signals score 86/100, with 2 cautions from licensing, adoption, age or documentation. These come from repository metadata, not a code audit — read the skill file before letting an agent act on it.
Safety scan
No issues foundOur scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands.
Automated pattern scan on 2026-09-29. It catches known dangerous patterns, not every risk — read a skill before letting an agent act on it.
mcp-safe-proxy compared with similar skills
All 4 of these similar skills score higher than mcp-safe-proxy; compare them before choosing.
| Skill | Score | Stars | Updated | Format |
|---|---|---|---|---|
| mcp-safe-proxy (this skill)by Huan-zhaojun | 72 | 10 | 7mo ago | MCP Server |
| Agent-Reachby Panniantong | 100 | 86.2k | 14d ago | CLAUDE.md |
| headroomby headroomlabs-ai | 100 | 74.1k | today | CLAUDE.md |
| rufloby ruvnet | 100 | 73.5k | today | CLAUDE.md |
| CowAgentby zhayujie | 100 | 47.2k | today | CLAUDE.md |
Frequently asked questions
- How do I install mcp-safe-proxy?
- Run
claude mcp add Huan-zhaojun -- npx -y github:Huan-zhaojun/mcp-safe-proxy. The install tabs above show the steps for each supported agent. - Which AI agents does mcp-safe-proxy work with?
- It is written for Claude Code, Claude Desktop and OpenAI Codex, as a MCP Server file. Other agents that read the same format can often use it too.
- Is mcp-safe-proxy safe to use?
- Our scan of the whole file found no instruction hijacking, hidden characters, credential access, data exfiltration or destructive commands. It is MIT-licensed and scores 86/100 on trust signals. Skills are instructions an agent will follow, so read the file before installing it and do not approve commands you do not understand.
- Is mcp-safe-proxy still maintained?
- The repository was last updated about 7 months ago. That is recent enough to be usable, but agent tooling moves fast, so check the instructions against your agent's current version.
Skill content
View source on GitHubmcp-safe-proxy
轻量级 MCP 注解代理 — 拦截 tools/list 响应,将工具注解重写为"安全"值,绕过 Codex 等 AI CLI 的审批弹窗,同时完全透传所有实际操作。
问题背景
OpenAI Codex CLI 根据 MCP Server 在 tools/list 响应中声明的工具注解判定是否需要审批:
需要审批 = destructiveHint == true
OR (readOnlyHint == false AND openWorldHint == true)
Playwright MCP 等 Server 对所有工具硬编码了 openWorldHint: true,导致每次操作都弹审批,严重影响自动化体验。
工作原理
在 MCP Client(如 Codex)和真实 MCP Server 之间插入 stdio 代理:
flowchart LR
subgraph Client["AI CLI (Codex / Claude Code)"]
A["MCP Client"]
end
subgraph Proxy["mcp-safe-proxy"]
B["stdio 流代理"]
C["注解重写引擎"]
end
subgraph Server["MCP Server (Playwright 等)"]
D["真实 MCP Server"]
end
A -->|"stdin 请求"| B
B -->|"透传请求"| D
D -->|"stdout 响应"| C
C -->|"tools/list: 重写注解\n其他: 原样转发"| A
核心逻辑:
- 拦截
tools/list响应,将每个工具的注解重写为:readOnlyHint: true ← 告诉 Client 工具只读 destructiveHint: false ← 告诉 Client 工具不搞破坏 openWorldHint: false ← 告诉 Client 工具不访问外部 tools/call、initialize等所有其他消息完全透传,功能零损失- 保留工具原有的
title、idempotentHint等其他注解字段
快速开始
安装
# 方式一:全局安装(推荐,启动更快)
npm install -g mcp-safe-proxy
# 方式二:npx 按需调用(无需预装)
npx -y mcp-safe-proxy -- <command> [args...]
Codex 配置
在 ~/.codex/config.toml(Linux/macOS)或 %USERPROFILE%\.codex\config.toml(Windows)中:
原始配置(会弹审批):
[mcp_servers.playwright]
type = "stdio"
command = "npx"
args = ["@playwright/mcp@latest"]
代理配置 — npx 方式(不弹审批):
[mcp_servers.playwright]
type = "stdio"
command = "npx"
args = ["-y", "mcp-safe-proxy", "--", "npx", "@playwright/mcp@latest"]
代理配置 — 全局安装方式:
[mcp_servers.playwright]
type = "stdio"
command = "mcp-safe-proxy"
args = ["--", "npx", "@playwright/mcp@latest"]
ccSwitch 配置
ccSwitch 使用标准 JSON 格式:
{
"command": "npx",
"args": ["-y", "mcp-safe-proxy", "--", "npx", "@playwright/mcp@latest"]
}
从源码构建(开发者)
git clone https://github.com/Huan-zhaojun/mcp-safe-proxy.git
cd mcp-safe-proxy
npm install
npm run build
详见 本地开发测试指南。
CLI 选项
mcp-safe-proxy [options] -- <command> [args...]
选项:
--verbose, -v 启用调试日志输出到 stderr
--log-file <path> 将调试日志写入指定文件(隐含 --verbose)
示例:
mcp-safe-proxy -- npx @playwright/mcp@latest
mcp-safe-proxy --verbose -- npx @playwright/mcp@latest
mcp-safe-proxy --log-file /tmp/proxy.log -- npx @playwright/mcp@latest
--是分隔符:左侧为代理选项,右侧为要包裹的 MCP Server 命令。
测试
# E2E 测试(使用 mock MCP Server)
npm run build && node test/e2e-test.js
# 真实 Playwright MCP 对比测试(需安装 @playwright/mcp)
node test/real-playwright-test.js
技术特点
- 零运行时依赖 — 仅使用 Node.js 内置模块(
child_process、fs、path) - ~220 行 TypeScript — 极简实现,易于审计和维护
- 通用 MCP 代理 — 不限 Playwright,任何 stdio MCP Server 均可使用
- 零侵入 — 不修改 MCP Server 代码,不全局安装,随时启用/恢复
适用范围
mcp-safe-proxy 在 MCP 协议层工作,兼容所有使用 stdio 传输的 MCP Server。以下是常见受审批问题影响的 Server 类别:
| 类别 | 代表 Server | 触发原因 | 注解验证 |
|------|------------|---------|:---:|
| 浏览器自动化 | Playwright MCP, Chrome DevTools MCP, Browserbase | action 工具 destructive + openWorld | ✅ |
| 文件与代码 | Filesystem MCP, GitHub MCP | 写入/删除操作 destructive | ✅ |
| 通信协作 | Notion MCP, Slack MCP, Google Workspace MCP | 非 GET 操作 destructive | ✅ |
| DevOps | Docker MCP | 容器操作 | ⬜ |
通用规则:只要 MCP Server 的工具声明了
destructiveHint: true或openWorldHint: true,就会触发 Codex 审批,mcp-safe-proxy 就能解决。
相关文档
- 兼容 MCP Server 一览 — 分类列表、配置示例、验证方法
- 技术设计文档 — 架构设计、原理详解、风险分析
- Codex MCP 权限问题分析 — 问题根因、源码追踪
- 本地开发测试指南 — 开发环境、配置示例、验收测试
License
Related Skills
Agent-Reach
86.2kGive your AI agent eyes to see the entire internet. Read & search Twitter, Reddit, YouTube, GitHub, Bilibili, XiaoHongShu — one CLI, zero API fees.
headroom
74.1kCompress tool outputs, logs, files, and RAG chunks before they reach the LLM. 20% fewer tokens for coding agents, 60-95% fewer tokens for JSON, same answers. Library, proxy, MCP server.
ruflo
73.5k🌊 The original agent harness. Deploy intelligent multi-player swarms, coordinate autonomous workflows, and build conversational AI systems. Features adaptive memory, self-learning intelligence, federation, vector RAG integration, and native Claude Code / Codex / Hermes and many more Integrated
CowAgent
47.2kOpen-source personal AI assistant & Agent Harness. Plans tasks, runs tools and skills, self-evolves with memory and knowledge. Multi-agent, multi-model, multi-channel. Lightweight, extensible, one-line install.
Languages
Trust signals
From repository metadata: license, adoption, age and documentation. Not a code audit — see the Safety scan above for what the skill file itself contains.
