Claude Code Plugins

Community-maintained marketplace

Feedback

mcp-server-development

@akiojin/llm-router
1
0

Expert guidance for building MCP (Model Context Protocol) servers using the TypeScript SDK. Use when developing MCP servers, implementing tools/resources/prompts, or working with the @modelcontextprotocol/sdk package. Covers server initialization, request handlers, Zod schemas, error handling, and JSON-RPC patterns.

Install Skill

1Download skill
2Enable skills in Claude

Open claude.ai/settings/capabilities and find the "Skills" section

3Upload to Claude

Click "Upload skill" and select the downloaded ZIP file

Note: Please verify skill by going through its instructions before using it.

SKILL.md

name mcp-server-development
description Expert guidance for building MCP (Model Context Protocol) servers using the TypeScript SDK. Use when developing MCP servers, implementing tools/resources/prompts, or working with the @modelcontextprotocol/sdk package. Covers server initialization, request handlers, Zod schemas, error handling, and JSON-RPC patterns.
allowed-tools Read, Grep, Glob, Bash, Edit, Write

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.

Hybrid Input (Compatibility)

Accept both Content-Length and NDJSON input for client compatibility:

  • Claude Desktop: Content-Length
  • Some CLI tools: NDJSON (newline-delimited)

Error Handling

Structured Errors

Use MCP error codes from the spec:

const McpError = {
  ParseError: -32700,
  InvalidRequest: -32600,
  MethodNotFound: -32601,
  InvalidParams: -32602,
  InternalError: -32603
};

throw new Error(JSON.stringify({
  code: McpError.InvalidParams,
  message: 'primitiveType must be one of: cube, sphere, cylinder'
}));

Error Response Format

{
  jsonrpc: '2.0',
  id: requestId,
  error: {
    code: -32602,
    message: 'Invalid params',
    data: { field: 'name', issue: 'required' }
  }
}

Unity-Specific Patterns

Command Protocol

Unity communication uses a simple request/response pattern:

const result = await this.unityConnection.send({
  command: 'gameobject_create',
  params: { name: 'Cube', primitiveType: 'cube' }
});

Workspace Root Resolution

ALWAYS include workspaceRoot in commands requiring file paths:

execute(params) {
  return this.unityConnection.send({
    command: 'screenshot_capture',
    workspaceRoot: config.workspaceRoot,
    ...params
  });
}

Testing Patterns

TDD for Handlers

  1. Contract test first: Define expected input/output schema
  2. Mock Unity connection: Isolate handler logic
  3. Test error paths: Invalid input, connection failures
  4. Test happy path last: After contracts are verified
describe('CreateGameObjectHandler', () => {
  it('validates primitiveType enum', async () => {
    const handler = new CreateGameObjectHandler(mockConnection);
    await assert.rejects(
      () => handler.handle({ name: 'Cube', primitiveType: 'invalid' }),
      /primitiveType must be one of/
    );
  });
});

Integration Testing

Test full request/response cycle including JSON-RPC framing:

const stdin = new PassThrough();
const stdout = new PassThrough();
const transport = new HybridStdioServerTransport(stdin, stdout);

stdin.write('Content-Length: 50\r\n\r\n{"jsonrpc":"2.0","id":1,"method":"ping"}');
// Assert stdout contains Content-Length response

Common Mistakes

Mixing framing formats:

  • NEVER: Output NDJSON after receiving Content-Length input
  • ALWAYS: Output Content-Length regardless of input format

Swallowing errors:

  • NEVER: catch (e) { return null; }
  • ALWAYS: Propagate errors with context

Missing validation:

  • NEVER: Trust raw input from clients
  • ALWAYS: Validate with Zod before processing

Blocking stdio:

  • NEVER: Synchronous operations on transport
  • ALWAYS: Use async/await throughout

Handler Registration

Register all handlers in a central index:

// src/handlers/index.js
export function createHandlers(unityConnection) {
  return [
    new SystemPingToolHandler(unityConnection),
    new CreateGameObjectHandler(unityConnection),
    new ScreenshotHandler(unityConnection),
    // ...
  ];
}

Remember

  • Content-Length always: Output framing must be consistent
  • Validate everything: Never trust client input
  • One handler, one tool: Keep handlers focused
  • Test error paths: Most bugs hide in error handling
  • Protocol compliance: Follow JSON-RPC 2.0 exactly