Skip to main content
Glama
Ailian0206

mcp-guardian

by Ailian0206

MCP Guardian

装进 Cursor / Codex 的本地 MCP 策略中间层:在 Agent 调用工具前 allow / deny / redact;高危在 同一次 Agent 对话 里批准。

仓库:https://github.com/Ailian0206/mcp-guardian
在线说明书:https://ailian0206.github.io/mcp-guardian/(GitHub Pages 静态站:介绍 / FAQ / Demo;不含 Dashboard API)
作品集:Evidence Graph → Work → MCP Guardian
决策文档:docs/product-redesign-2026-07-16.md

主路径(请按这个理解)

  1. bash scripts/install.sh → 重启 IDE

  2. Agent 正常用工具;多数策略自动生效

  3. 高危 → 返回 approval_required → Agent 问你 → 你确认 → Agent 调 guardian_decide

  4. 不需要打开网页审批,也不需要另开终端跑 approvals

Web 站点只做介绍 / FAQ / 可选策略试跑。

Related MCP server: evav-gateway

一键安装

git clone https://github.com/Ailian0206/mcp-guardian.git
cd mcp-guardian
bash scripts/install.sh

写入 Cursor ~/.cursor/mcp.json、Codex ~/.codex/config.toml 与 ~/.mcp-guardian/。
仅一边:bash scripts/install.sh --cursor 或 --codex。

真实下游(官方 Filesystem MCP)

默认安装是演示下游。接真实磁盘目录:

pnpm build
node packages/gateway/dist/cli.js install --cursor --profile filesystem --workspace /ABS/PATH/TO/DIR
# Reload MCP 后:读放行;写/改/移 → 会话内 confirm_code + guardian_decide

切回演示:--profile demos。示例配置:examples/real-filesystem.config.yaml。

Agent 会话内审批

当策略为 require_approval 时,工具结果包含:

  • approval_id

  • agent_instructions:先问用户,再调用 guardian_decide

同 MCP 还暴露:

  • guardian_pending — 查看待批

  • guardian_decide — { id, decision: "allow" | "deny", confirm_code? };allow 须带码

Web(说明书 / 作品集门面)

在线:https://ailian0206.github.io/mcp-guardian/

pnpm dev:web   # http://127.0.0.1:3040  → /  /faq  /demo
pnpm build:pages  # 静态导出 → apps/web/out(供 GitHub Pages)
  • /:价值、安装步骤、日常流程、真实下游、与 Trace 差异、FAQ 摘要、演示录屏

  • /faq:入门 / 日常 / 真实下游 / 排错

  • /demo:浏览器内现场试跑策略 + 只读审计回放(不是审批台)

  • /app:可选 Dashboard(仅本地 / 非 Pages;需服务端 API)

开发验收

pnpm install && pnpm build
pnpm lint && pnpm typecheck && pnpm test
bash scenarios/a1-a8.sh
bash scenarios/ide-smoke.sh
bash scenarios/real-filesystem.sh

License

MIT

Related MCP Connectors

Related MCP Servers

  • A
    license
    Not graded
    quality
    D
    maintenance
    A policy-enforcing MCP gateway that intercepts all tool calls to downstream MCP servers, applying allow/deny/ask rules with human approval and audit logging for safe access to dangerous tools.
    11 npm
    MIT
  • A
    license
    Not graded
    quality
    D
    maintenance
    Governed MCP gateway that lets AI agents call tools with policy enforcement, prompt-injection screening, a kill-switch, and tamper-evident signed audit logs.
    Apache 2.0
  • F
    license
    Not graded
    quality
    D
    maintenance
    MCP server that provides a security gateway for AI agents, enforcing allow/confirm/deny policies on tool calls and requiring human approval for risky operations, with full audit logging.
    -
  • A
    license
    Not graded
    quality
    C
    maintenance
    Acts as an MCP gateway between AI agents and upstream tool servers, evaluating YAML policies on every tool call, injecting real credentials at the last mile, scanning requests and responses for secrets, PII, and prompt injection, and recording each action in a tamper-evident audit log. Also supports HTTP forward-proxy governance with rate limiting, cost tracking, and policy replay against historical audit data.
    4
    Apache 2.0