首页龙虾技能列表 › MCP Vercel — MCP工具

MCP Vercel — MCP工具

v1.0.0

[AI辅助] Deploy a remote MCP server on Vercel with Next.js and mcp-handler. Use this skill whenever the user wants to create an MCP server, deploy MCP to Vercel, set...

0· 132·0 当前·0 累计
by @lucaperret (Luca)·MIT-0
下载技能包
License
MIT-0
最后更新
2026/3/20
安全扫描
VirusTotal
无害
查看报告
OpenClaw
安全
high confidence
The skill's requirements and runtime instructions are consistent with its stated purpose (deploying an MCP server on Vercel using Next.js and mcp-handler); nothing in the package attempts to access unrelated credentials or perform unexpected installs.
评估建议
This skill appears coherent for adding an MCP endpoint on Vercel, but before installing/applying it: (1) review the npm packages (mcp-handler, @modelcontextprotocol/sdk, zod) for trustworthiness and recent releases; (2) when registering tools, avoid implementing handlers that perform sensitive data reads or expose secrets — test in a staging project first; (3) deploying to Vercel requires your Vercel account/CLI auth — keep tokens secure and do not paste them into third-party UIs; (4) follow the...
详细分析 ▾
用途与能力
The name/description match the instructions: the SKILL.md shows how to add a Next.js route using mcp-handler and @modelcontextprotocol/sdk and how to deploy to Vercel. It does not request unrelated binaries, credentials, or config paths.
指令范围
Instructions are focused on creating an MCP route, registering tools, and Vercel deployment. They ask the developer to run npm install and to copy compiled artifacts into the Vercel root if using a root directory. That file-copy guidance implies reading project build outputs (e.g., ../../dist) — this is expected for build/deploy steps but you should ensure those paths do not accidentally reference or expose secrets or unrelated files.
安装机制
This is an instruction-only skill (no install spec). The SKILL.md tells the user to run npm install for public packages (mcp-handler, @modelcontextprotocol/sdk, zod). There are no downloads from arbitrary URLs or automatic install steps performed by the skill itself.
凭证需求
The skill declares no required environment variables or credentials. In practice, deploying with the Vercel CLI requires Vercel auth (not requested by this skill) and you may add OAuth tokens for protected tools — the skill does not demand unrelated secrets.
持久化与权限
always is false, the skill is user-invokable and not force-included. There is no instruction to modify other skills or global agent settings.
安全有层次,运行前请审查代码。

License

MIT-0

可自由使用、修改和再分发,无需署名。

运行时依赖

无特殊依赖

版本

latestv1.0.02026/3/20

Initial release of mcp-vercel. - Guides users to deploy a remote MCP server on Vercel using Next.js and mcp-handler. - Provides setup instructions, including tool registration and deployment steps. - Explains correct routing, parameter annotation, and safety best practices for MCP tools. - Details deployment caveats specific to Vercel’s serverless environment. - Includes tips for authentication and publishing to Smithery.

● 无害

安装命令 点击复制

官方npx clawhub@latest install mcp-vercel
镜像加速npx clawhub@latest install mcp-vercel --registry https://cn.clawhub-mirror.com

技能文档

Create a production-ready remote MCP server on Vercel using Next.js and mcp-handler. The server communicates via Streamable HTTP and works with Claude Desktop, claude.ai, Smithery, and any MCP client.

为什么 approach

Vercel's serverless functions are ideal for MCP servers because MCP's Streamable HTTP transport is stateless — each request is independent, which maps perfectly to serverless. No persistent connections needed. The mcp-handler package from Vercel handles all the protocol details.

Quick setup

1. Install dependencies

npm install mcp-handler @modelcontextprotocol/sdk zod

2. 创建 MCP 路由

Create app/api/mcp/route.ts:

import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
import { createMcpHandler } from 'mcp-handler';
import { z } from 'zod';

const handler = createMcpHandler( (server: McpServer) => { // Register your tools here server.tool( 'example_tool', 'What this tool does — be specific', { query: z.string().describe('What the parameter is for') }, { readOnlyHint: true, destructiveHint: false, title: 'Example Tool' }, async ({ query }) => ({ content: [{ type: 'text', text: Result for: ${query} }], }), ); }, { serverInfo: { name: 'my-server', version: '1.0.0' }, }, { streamableHttpEndpoint: '/api/mcp', maxDuration: 60, }, );

export { handler as GET, handler as POST, handler as DELETE };

3. Deploy 和 test

vercel deploy --prod

# Verify curl -X POST https://your-app.vercel.app/api/mcp \ -H "Content-Type: application/json" \ -H "Accept: application/json, text/event-stream" \ -d '{"jsonrpc":"2.0","method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}},"id":1}'

You should get back serverInfo with your server name and version.

Tool design

Safety annotations (必填)

Every tool must have annotations. MCP clients use these to decide how cautiously to invoke tools.

// Read-only (search, get, list, fetch)
{ readOnlyHint: true, destructiveHint: false, title: 'Search Items' }

// Write but not destructive (create, add, update) { readOnlyHint: false, destructiveHint: false, title: 'Create Item' }

// Destructive (delete, remove, overwrite) { readOnlyHint: false, destructiveHint: true, title: 'Delete Item' }

Parameter descriptions

Every parameter needs a .describe() — this is how MCP clients know what to pass.

{
  query: z.string().describe('Search query text'),
  limit: z.number().optional().default(10).describe('Max results to return'),
  type: z.enum(['artist', 'album', 'track']).describe('Type of content'),
}

MCP prompts (可选 但是 recommended)

Prompt templates improve discoverability on Smithery and give users ready-made starting points.

server.prompt('find_items', 'Search for items by name',
  { name: z.string().describe('Item name') },
  ({ name }) => ({
    messages: [{
      role: 'user' as const,
      content: { type: 'text' as const, text: Find ${name} and show details },
    }],
  }),
);

Routing — streamableHttpEndpoint gotcha

Use streamableHttpEndpoint, NOT basePath:

// CORRECT — endpoint at /api/mcp
{ streamableHttpEndpoint: '/api/mcp' }

// WRONG — creates endpoint at /api/mcp/mcp (doubled path) { basePath: '/api/mcp' }

The basePath option appends /mcp to whatever you give it. Since your route file is already at app/api/mcp/route.ts, that creates /api/mcp/mcp.

Vercel deployment pitfalls

Root Directory isolation

If your Vercel project uses a Root Directory (like site/), the deployed function CANNOT access files outside that directory. This means import from '../../dist/' will fail at runtime even if it compiles locally.

Solution: 复制 compiled files 进入 site directory 和 commit them. 使用 prebuild script 到 keep them 在...中 同步:

// scripts/copy-deps.js
const fs = require('fs');
const path = require('path');
const src = path.resolve(__dirname, '../../dist');
const dest = path.resolve(__dirname, '../lib/deps');
fs.mkdirSync(dest, { recursive: true });
for (const file of fs.readdirSync(src)) {
  if (file.endsWith('.js') || file.endsWith('.d.ts')) {
    fs.copyFileSync(path.join(src, file), path.join(dest, file));
  }
}

Add to package.json: "prebuild": "node scripts/copy-deps.js"

Serverless 读取-仅 filesystem

Vercel functions run on a read-only filesystem with no home directory. If your code writes files (sessions, temp data), wrap in try/catch:

try {
  fs.mkdirSync(dir, { recursive: true });
  fs.writeFileSync(filepath, data);
} catch {
  // Serverless environment — skip filesystem writes
}

Turbopack CJS/ESM mismatch

If importing CommonJS .js files while the parent package.json doesn't have "type": "module", Turbopack will error. Solution: import from compiled .js files bundled within the site directory, not from TypeScript source files in the parent.

Adding authentication

For OAuth-protected servers, see the mcp-oauth skill which covers the complete OAuth 2.0 PKCE flow with withMcpAuth, including dynamic client registration and token storage.

Publishing 到 Smithery

After deploying, publish to Smithery for broader distribution:

  • Go 到 https://smithery.ai/新的
  • Enter MCP server URL
  • Choose namespace/server-id
  • Smithery scans tools automatically

If your server requires auth, Smithery will prompt you to connect during scanning.

Reference

数据来源:ClawHub ↗ · 中文优化:龙虾技能库
OpenClaw 技能定制 / 插件定制 / 私有工作流定制

免费技能或插件可能存在安全风险,如需更匹配、更安全的方案,建议联系付费定制

了解定制服务