如何在 AI Agents 中正确使用 MCP 服务器

Philipp Schmid 2026-04-28 8.9 KB 56.8K 216 𝕏 文章列表

文章讨论了 MCP 服务器在 AI Agents 中的应用。盲目启用 MCP 会导致上下文膨胀、成本增加和性能下降。文章提出了两种有效的使用模式：一是显式 MCP 服务器（内联工具注入），通过 @mention 按需加载工具，适合用户驱动的临时需求；二是子代理 MCP 服务器，将 MCP 服务器声明在子代理定义中，利用 allowed_tools 进行最小权限范围限定，适合代码审查或支持代理等特定场景。

# How to correctly use MCP servers with your AI Agents **作者**: Philipp Schmid **日期**: 2026-04-27T15:08:35.000Z **来源**: [https://x.com/_philschmid/status/2048781360643375434](https://x.com/_philschmid/status/2048781360643375434) --- ![Image](/doc/1162/image/img_1be89fce9823.jpg) MCP servers are not dead. But blindly enabling them bloats your context, which leads to higher cost and worse performance. Unlike Agent Skills, MCP servers don't come with progressive disclosure out of the box. It is your responsibility to select the tools needed for the task at hand. Here are two proven patterns on how to correctly use MCP servers and avoid the bloat. ## 1. Explicit MCP Servers: Inline Tool Injection MCP servers stay opt-in. Servers are referenced in your prompt with an @mention, and the agent resolves, fetches, and injects the tools before the model request. ![Image](/doc/1162/image/img_ebbe48bb91e2.jpg) Writing @github or @slack in a prompt triggers the agent to: 1. Resolve the @mention to a registered MCP server URL 2. Fetch tool schemas from the server 3. Inject them into the tools[] array of the API request 4. Forward everything to the model, which decides what to call Nothing loads unless requested. The tool surface stays small. ``` // Pseudo-code async function handlePrompt(prompt: string) { // Detect @mentions in the prompt const mentions = parseMentions(prompt); // ["@github", "@slack"] // Resolve each mention to an MCP server const servers = mentions.map(m => mcpServers.resolve(m)); // Fetch tool schemas from each server const mcpTools = await Promise.all( servers.map(s => s.listTools()) ); // Inject into the API request alongside native tools const response = await llmCall({ prompt, tools: [...nativeTools, ...mcpTools], }); // handle response, call tools, etc ... } ``` When to use this: MCP usage is occasional and user-driven. Someone asks a question that needs Slack or GitHub data, they @mention the server, and it gets pulled in for that request only. Keeps costs low and avoids tool noise on tasks that don't need external integrations. ## 2. Subagent MCP Servers MCP servers are declared in subagents definition and automatically available at runtime alongside native tools like read_file or bash. ![Image](/doc/1162/image/img_c06efb38f220.jpg) Each subagent has its own configuration that declares its model, tools,... and optional MCP servers. You can scope MCP servers to specific tools using allowed_tools. ``` # Pseudo-code: code_reviewer.md --- name: code-reviewer model: gemini-3-flash mcp_servers: - url: https://github-mcp.example allowed_tools: - list_pulls - list_reviews - get_diff --- You are a code reviewer. Review open PRs... ``` The main agent registers subagents. When invoked, each subagent connects to its declared MCP servers and merges those tools with native ones. ``` // Pseudo-code // Register subagents as tools for the orchestrator const codeReviewer = createSubagent("./agents/code_reviewer.md"); const slackMonitor = createSubagent("./agents/slack_monitor.md"); async function handlePrompt(prompt: string) { // pre-processing ... const response = await llmCall({ prompt, tools: [...nativeTools, codeReviewer, slackMonitor], }); // handle response, call tools, etc ... } ``` allowed_tools gives you least-privilege scoping without forking the MCP server. When to use this: The use case dictates the tools. A code review agent always needs GitHub, a support agent always needs Zendesk. The MCP servers are part of what the agent is, not something a user opts into per request. allowed_tools keeps each agent scoped to what it actually needs. ## 相关链接 - [Philipp Schmid](https://x.com/_philschmid) - [@_philschmid](https://x.com/_philschmid) - [24K](https://x.com/_philschmid/status/2048781360643375434/analytics) - [@mention](https://x.com/@mention) - [@github](https://x.com/@github) - [@slack](https://x.com/@slack) - [@mention](https://x.com/@mention) - [@mention](https://x.com/@mention) - [Upgrade to Premium](https://x.com/i/premium_sign_up) - [11:08 PM · Apr 27, 2026](https://x.com/_philschmid/status/2048781360643375434) - [24.7K Views](https://x.com/_philschmid/status/2048781360643375434/analytics) - [View quotes](https://x.com/_philschmid/status/2048781360643375434/quotes) --- *导出时间: 2026/4/28 08:46:04* --- ## 中文翻译 # 如何在 AI Agent 中正确使用 MCP 服务器 **作者**： Philipp Schmid **日期**： 2026-04-27T15:08:35.000Z **来源**： [https://x.com/_philschmid/status/2048781360643375434](https://x.com/_philschmid/status/2048781360643375434) --- ![Image](/doc/1162/image/img_1be89fce9823.jpg) MCP 服务器并没有死。但盲目启用它们会让你的上下文膨胀，导致更高的成本和更差的性能。与 Agent Skills 不同，MCP 服务器开箱即用时不带渐进式披露机制。你需要自己负责为当前任务选择所需的工具。下面是两个经过验证的模式，教你如何正确使用 MCP 服务器并避免上下文膨胀。 ## 1. 显式 MCP 服务器：内联工具注入 MCP 服务器保持按需启用。通过在提示词中使用 @mention 来引用服务器，Agent 会在模型请求之前解析、获取并注入工具。 ![Image](/doc/1162/image/img_ebbe48bb91e2.jpg) 在提示词中写入 @github 或 @slack 会触发 Agent： 1. 将 @mention 解析为已注册的 MCP 服务器 URL 2. 从服务器获取工具 schema 3. 将其注入到 API 请求的 tools[] 数组中 4. 将所有内容转发给模型，由模型决定调用什么除非被请求，否则不会加载任何内容。工具面保持很小。 ``` // Pseudo-code async function handlePrompt(prompt: string) { // Detect @mentions in the prompt const mentions = parseMentions(prompt); // ["@github", "@slack"] // Resolve each mention to an MCP server const servers = mentions.map(m => mcpServers.resolve(m)); // Fetch tool schemas from each server const mcpTools = await Promise.all( servers.map(s => s.listTools()) ); // Inject into the API request alongside native tools const response = await llmCall({ prompt, tools: [...nativeTools, ...mcpTools], }); // handle response, call tools, etc ... } ``` 何时使用：MCP 的使用是偶尔且由用户驱动的。当有人提出需要 Slack 或 GitHub 数据的问题时，他们 @mention 对应的服务器，该服务器仅在此次请求中被引入。这样能保持低成本，并避免在不需要外部集成的任务上产生工具噪音。 ## 2. 子代理 MCP 服务器 MCP 服务器在子代理（subagent）的定义中声明，并在运行时自动与 read_file 或 bash 等原生工具一起可用。 ![Image](/doc/1162/image/img_c06efb38f220.jpg) 每个子代理都有自己的配置，声明其模型、工具……以及可选的 MCP 服务器。你可以使用 allowed_tools 将 MCP 服务器限定到特定工具。 ``` # Pseudo-code: code_reviewer.md --- name: code-reviewer model: gemini-3-flash mcp_servers: - url: https://github-mcp.example allowed_tools: - list_pulls - list_reviews - get_diff --- You are a code reviewer. Review open PRs... ``` 主代理注册子代理。当被调用时，每个子代理会连接到其声明的 MCP 服务器，并将这些工具与原生工具合并。 ``` // Pseudo-code // Register subagents as tools for the orchestrator const codeReviewer = createSubagent("./agents/code_reviewer.md"); const slackMonitor = createSubagent("./agents/slack_monitor.md"); async function handlePrompt(prompt: string) { // pre-processing ... const response = await llmCall({ prompt, tools: [...nativeTools, codeReviewer, slackMonitor], }); // handle response, call tools, etc ... } ``` allowed_tools 让你无需 fork MCP 服务器即可实现最小权限的范围限定。何时使用：由用例决定所需工具。代码审查代理总是需要 GitHub，客服代理总是需要 Zendesk。MCP 服务器是代理本身的一部分，而不是用户在每次请求中自行选择加入的东西。allowed_tools 使每个代理的范围仅限于它实际需要的内容。 ## 相关链接 - [Philipp Schmid](https://x.com/_philschmid) - [@_philschmid](https://x.com/_philschmid) - [24K](https://x.com/_philschmid/status/2048781360643375434/analytics) - [@mention](https://x.com/@mention) - [@github](https://x.com/@github) - [@slack](https://x.com/@slack) - [@mention](https://x.com/@mention) - [@mention](https://x.com/@mention) - [升级到 Premium](https://x.com/i/premium_sign_up) - [2026年4月27日晚上11:08](https://x.com/_philschmid/status/2048781360643375434) - [24.7K 次浏览](https://x.com/_philschmid/status/2048781360643375434/analytics) - [查看引用](https://x.com/_philschmid/status/2048781360643375434/quotes) --- *导出时间: 2026/4/28 08:46:04*