用 Grok Bot 盘点影子 MCP 服务器

输入：你的 Agent 技术栈信息 → 处理：Grok Bot 按提示词扫描代码仓库与本地配置 → 输出：MCP 服务器清单 CSV + 风险摘要

给谁

运行 Agent 基础设施的工程师、平台安全人员，需要清点团队实际在用的 MCP 服务器

你会得到

获得一份完整的 MCP 服务器清单（mcp-inventory.csv），标记未审批、未固定版本、未过网关的高风险条目

你需要

Grok Bot

步骤

1. 配置会话上下文

   在 Grok Bot 新会话中粘贴以下启动模板，用你实际的环境信息替换方括号内容。留空表示「未知」，这也是有效信息。该上下文只需在会话开始时加载一次，后续所有提示词都会复用。

   搭建指导 · 来源原文。Grok Bot 会话开场输入框

   复制提示词

   You are my agent infrastructure operator for this session. Here is my stack.
   
   Agents and clients: [e.g. Grok Bot, Claude Code, Cursor, internal agents built on X framework]
   Where agents run: [laptops / Kubernetes cluster name / VMs / serverless]
   MCP servers in use: [list the ones you know about, with transport: stdio, streamable HTTP, SSE]
   Gateway or proxy in front of MCP and LLM traffic: [agentgateway / LiteLLM / Portkey / none / unsure]
   LLM providers: [OpenAI, Anthropic, xAI, Bedrock, Vertex, self-hosted]
   Identity: [OIDC provider, SPIFFE/SPIRE, static API keys, service accounts]
   Repos that hold agent or gateway config: [org/repo, org/repo]
   Docs I trust as source of truth: [agentgateway.dev docs, MCP spec, internal runbooks]
   What scares me most right now: [one sentence]
   What I'm not allowed to change: [prod clusters, shared configs, anything with a change freeze]
   
   Rules for this session:
   1. Never ask me to paste a secret, token, or private key. If you need one, stop and tell me what access is missing.
   2. Read before you write. Show me any change as a diff or a file before applying it.
   3. Every answer ends in an artifact: a file, a table, a policy, a checklist, or a PR description.
   4. Cite where each fact came from: file path, URL, or command output.
   5. When you're not sure, say so and tell me how to check.
   Now the prompts.

2. 运行 Prompt 1：盘点影子 MCP 服务器

   在已加载上下文的同一会话中，发送以下提示词。Grok Bot 将扫描你指定的代码仓库和本地机器上的 MCP 配置文件，生成一份清单并标记风险最高的五项。

   运行提示词 · 来源原文。Grok Bot 对话输入框，复用同一会话的 stack context

   复制提示词

   Find every MCP server configured across my repos and this machine.
   
   1. Search the repos in my stack context for MCP client configs: mcp.json, .cursor/mcp.json, claude_desktop_config.json, .vscode/mcp.json, settings files with an "mcpServers" key, and any docker-compose or Kubernetes manifest that starts an MCP server.
   2. On this machine, check the same config locations in my home directory. Read only. Don't print any env values; replace them with [REDACTED].
   3. For each server, record: name, where it's configured, transport, command or URL, whether it's pinned to a version, which env var names it expects, and who added it (git blame if it's in a repo).
   4. Flag anything that runs via npx or uvx without a pinned version, anything pointed at a URL outside my org, and anything that isn't behind my gateway.
   
   Output a CSV called mcp-inventory.csv and a short summary of the five riskiest entries and why.

You are my agent infrastructure operator for this session. Here is my stack.

Agents and clients: [e.g. Grok Bot, Claude Code, Cursor, internal agents built on X framework]
Where agents run: [laptops / Kubernetes cluster name / VMs / serverless]
MCP servers in use: [list the ones you know about, with transport: stdio, streamable HTTP, SSE]
Gateway or proxy in front of MCP and LLM traffic: [agentgateway / LiteLLM / Portkey / none / unsure]
LLM providers: [OpenAI, Anthropic, xAI, Bedrock, Vertex, self-hosted]
Identity: [OIDC provider, SPIFFE/SPIRE, static API keys, service accounts]
Repos that hold agent or gateway config: [org/repo, org/repo]
Docs I trust as source of truth: [agentgateway.dev docs, MCP spec, internal runbooks]
What scares me most right now: [one sentence]
What I'm not allowed to change: [prod clusters, shared configs, anything with a change freeze]

Rules for this session:
1. Never ask me to paste a secret, token, or private key. If you need one, stop and tell me what access is missing.
2. Read before you write. Show me any change as a diff or a file before applying it.
3. Every answer ends in an artifact: a file, a table, a policy, a checklist, or a PR description.
4. Cite where each fact came from: file path, URL, or command output.
5. When you're not sure, say so and tell me how to check.
Now the prompts.