mcp-server-builder

When to Use

Trigger this skill when building custom tool servers for LLM Agents, wrapping APIs as MCP tools, or debugging MCP connections.

Core Concepts

MCP Architecture

• Host: the client running the LLM (Claude Desktop, IDE, etc.)
• Client: the middleware connecting to Servers
• Server: exposes Tools, Resources, and Prompts

Tool Definition

server.tool("query_database",
  { sql: z.string().describe("SQL query statement") },
  async ({ sql }) => {
    const result = await db.query(sql)
    return { content: [{ type: "text", text: JSON.stringify(result) }] }
  }
)

Resource Exposure

• Static resources: server.resource() registers fixed URIs
• Dynamic resources: template URI {id} matches parameters
• Subscription: notify clients on resource changes

Transport

• stdio: local process communication, simplest
• SSE: remote service, HTTP long connection
• Rule: stdio for local, SSE for remote

Security

• Validate all tool inputs with Zod schemas
• Add confirmation prompts for dangerous operations (annotations.destructiveHint)
• Load sensitive data on-demand via resources, never in prompts