FAQ

General

What is MCP?

The Model Context Protocol (MCP) is an open standard that lets AI applications discover and call functions on your server — like a universal plugin system for AI. It's developed by Anthropic and supported by Claude, Cursor, VS Code Copilot, and more.

How is this different from building a REST API?

A REST API requires you to define endpoints, handle HTTP parsing, manage auth, and write client SDKs. MCP servers use a standard protocol (JSON-RPC) so any MCP client can discover and call your tools automatically — no per-client SDK needed.

Do I need to know JSON-RPC?

No. MCPCraft handles all JSON-RPC serialization under the hood. You define plain TypeScript objects and functions.

Is this production ready?

MCPCraft is in active development. The core API is stable, but treat it as you would any early-stage tool — test thoroughly before deploying to production.

Compatibility

Which clients work with MCPCraft servers?

Any MCP-compatible client. This includes:

• Claude Desktop (Anthropic)
• Claude Code (Anthropic)
• Cursor (via .cursor/rules)
• VS Code Copilot (via GitHub Copilot MCP)
• Windsurf (Codeium)
• Continue.dev (open source)
• Custom clients built with @modelcontextprotocol/sdk

Does this work with ChatGPT?

ChatGPT supports the MCP protocol through its plugin system and can connect to MCP servers.

Can I use this with the OpenAI API directly?

Yes. The OpenAI Assistants API and function calling can consume MCP tools by using the auto-generated JSON schema output — no SDK abstraction needed.

Technical

What transport protocols are supported?

• stdio — local/desktop use (Claude Desktop, VS Code)
• SSE — remote servers (Vercel, Docker, VPS)

What Node.js version is required?

Node.js 18 or later. TypeScript 5+ recommended.

Can I use this with Deno or Bun?

The SDK targets Node.js runtime. Deno and Bun compatibility via Node compat layer is experimental.

How do I handle errors?

Throw from your run handler — MCPCraft catches it and returns a proper JSON-RPC error response to the client.

run: async ({ userId }) => {
  const user = await db.findUser(userId)
  if (!user) throw new Error("User not found")
  return user
}

Can I return binary data?

MCP supports text and blob content types. Use { type: "blob", data: base64String, mimeType: "image/png" }.

Deployment

Can I deploy to Vercel?

Yes. See the Deployment guide for Vercel serverless setup.

Can I deploy to Docker?

Yes. See the Deployment guide for Docker configuration.

Can I deploy to Railway or Render?

Yes. Both support Node.js environments with environment variable configuration.

Troubleshooting

My server starts but the client can't find tools

Make sure you're calling server.add(tool(...)) before server.start(). Tools registered after start() won't be available.

Claude Desktop shows "No tools found"

Check that:

1. Your server builds without errors (npm run build)
2. The path in claude_desktop_config.json is correct
3. The server file exports a running server

The inspector shows "Connection refused"

If using stdio, the server runs as a child process — you connect to the process, not a port. Use the inspector command with your server file:

npx @modelcontextprotocol/inspector node dist/server.js

If using SSE, ensure your server is running and the URL is accessible.