MCP Builder

Build MCP (Model Context Protocol) servers that give Claude and other LLMs direct access to your APIs, databases, and external tools.

What This Skill Does

The Challenge: LLMs are powerful but isolated. Connecting them to your CRM, analytics platform, or internal tools requires custom integrations for each model and use case.

The Solution: MCP Builder skill scaffolds MCP servers that expose any external service as standardized tools LLMs can call. Supports HTTP APIs, databases, file systems, and custom logic with proper authentication and error handling.

Activation

Implicit: Activates automatically when creating MCP server integrations or building LLM tool integrations.

Explicit: Activate via prompt:

Activate mcp-builder skill to create an MCP server for our Stripe billing API

Capabilities

1. MCP Server Scaffolding

Generate production-ready MCP server boilerplate with TypeScript.

Project structure:

mcp-server/
├── src/
│   ├── index.ts          # Server entry point
│   ├── tools/            # Tool definitions
│   └── resources/        # Resource definitions
├── package.json
└── README.md

2. Tool and Resource Definitions

Define tools LLMs can call and resources they can read.

Tool definition pattern:

server.setRequestHandler(CallToolRequestSchema, async (request) => {
  const { name, arguments: args } = request.params;
  // Handle tool calls
});

Resource types: REST API endpoints, database queries, file reads, webhooks

3. Authentication Patterns

Secure MCP servers with API keys, OAuth, or JWT.

Auth methods supported:

Method	Use Case
API Key (header)	Internal tools, simple integrations
OAuth 2.0	Third-party services (Google, GitHub)
JWT	Multi-tenant SaaS integrations
Bearer token	Most REST APIs

Prerequisites

Node.js 18+ and TypeScript
Target API credentials and documentation
@modelcontextprotocol/sdk package

Configuration

Install MCP SDK:

npm install @modelcontextprotocol/sdk

Environment variables (.env):

MCP_SERVER_PORT=3000
TARGET_API_KEY=your_api_key
TARGET_API_BASE_URL=https://api.example.com

Best Practices

1. One responsibility per tool Each MCP tool should do one thing. Avoid combining multiple API calls into one tool.

2. Schema-validate all inputs Use Zod schemas for tool input validation. LLMs pass unexpected values.

3. Return structured, parseable output LLMs work better with consistent JSON schemas than freeform text responses.

Common Use Cases

Use Case 1: CRM Integration MCP Server

Scenario: Give Claude direct access to HubSpot contacts and deals.

Tools to expose:

search_contacts - Search by name, email, company
get_deal_pipeline - Retrieve deal stages and values
create_note - Add activity notes to contacts
update_deal_stage - Move deals through pipeline

Output: Functional MCP server deployable via Claude Desktop config.

Use Case 2: Analytics Dashboard MCP Server

Scenario: Let marketing agents query GA4 data directly.

Tools to expose:

get_traffic_report - Sessions, users by date range
get_conversion_data - Goals and conversion rates
compare_periods - YoY/MoM performance

Output: MCP server with GA4 tool access for conversational analytics.

Troubleshooting

Issue: Tool calls return malformed JSON Solution: Add explicit JSON.stringify and schema validation in response handlers.

Issue: Authentication fails on API calls Solution: Verify credentials are loaded from environment, not hardcoded. Check token expiry.

Issue: Claude doesn’t discover tools Solution: Ensure server is registered in Claude Desktop claude_desktop_config.json.

MCP Management - Manage and use existing MCP servers
Web Frameworks - Backend API development

/ckm:mcp-builder - Scaffold new MCP server
/ckm:mcp-management - Discover and use MCP servers