llmlb: Skill for Claude Code

This skill provides comprehensive guidance for building robust MCP servers, with specific focus on the unity-mcp-server architecture and patterns.

Core Philosophy

MCP servers bridge AI assistants to external systems. They must be:

• Reliable: Handle errors gracefully, never crash unexpectedly
• Discoverable: Tools should have clear, self-documenting schemas
• Performant: Minimize latency, especially for stdio transport
• Protocol-compliant: Follow JSON-RPC 2.0 and MCP spec exactly

Architecture Patterns

Handler-Based Design

ALWAYS use a handler class per tool. Each handler encapsulates:

• Input validation (Zod schema)
• Business logic execution
• Error handling and response formatting

// Pattern: One handler per tool
export class SystemPingToolHandler extends BaseToolHandler {
  constructor(unityConnection) {
    super({
      name: 'system_ping',
      description: 'Check Unity Editor connectivity',
      inputSchema: { type: 'object', properties: {} }
    });
    this.unityConnection = unityConnection;
  }

  async execute(params) {
    const result = await this.unityConnection.send({ command: 'ping' });
    return { status: 'ok', ...result };
  }
}

BaseToolHandler Contract

All handlers MUST:

1. Call super() with tool metadata
2. Implement execute(params) method
3. Return plain objects (framework handles JSON-RPC wrapping)
4. Throw errors with descriptive messages

Input Validation with Zod

ALWAYS validate inputs before processing:

import { z } from 'zod';

const inputSchema = z.object({
  name: z.string().min(1).describe('GameObject name'),
  primitiveType: z.enum(['cube', 'sphere', 'cylinder']).optional()
});

validate(input) {
  return inputSchema.parse(input);
}

Transport Layer

Content-Length Framing (Standard)

MCP uses LSP-style Content-Length framing for stdio:

Content-Length: 123\r\n
\r\n
{"jsonrpc":"2.0","id":1,"method":"tools/call"...}

ALWAYS use Content-Length for output. NEVER mix framing formats in a session.