MCP Server Authoring
What MCP Servers Do
MCP servers expose tools, resources, and prompts to Claude via the Model Context Protocol. They run as separate processes (stdio) or HTTP endpoints that Claude communicates with.
Project Structure (dev-suite)
mcp-servers/{server-name}/
├── package.json # @dev-suite/{name}, main: dist/index.js
├── tsconfig.json # TypeScript config
├── metadata.json # Server metadata for dev-suite dashboard
└── src/
└── index.ts # Server implementation
TypeScript MCP Server Template
See quick-ref/typescript-template.md for the complete starter template.
import { McpServer } from '@modelcontextprotocol/sdk/server/mcp.js';
import { StdioServerTransport } from '@modelcontextprotocol/sdk/server/stdio.js';
import { z } from 'zod';
const server = new McpServer({
name: 'my-server',
version: '1.0.0',
});
// Define a tool
server.tool(
'search_docs', // Tool name
'Search documentation by query', // Description
{ query: z.string(), limit: z.number().optional().default(10) }, // Input schema
async ({ query, limit }) => { // Handler
const results = await searchDocumentation(query, limit);
return {
content: [{ type: 'text', text: JSON.stringify(results, null, 2) }],
};
}
);
// Start server
const transport = new StdioServerTransport();
await server.connect(transport);
Tool Definition Best Practices
- Name: snake_case, descriptive (
search_docsnotsearch) - Description: clear, includes when to use it
- Input schema: Zod validation with sensible defaults
- Output: Always return
contentarray withtype: 'text' - Errors: Throw errors with descriptive messages — MCP SDK handles formatting
metadata.json (dev-suite specific)
See quick-ref/metadata-schema.md for the complete schema.
{
"name": "my-server",
"description": "Full description of what this server does",
"shortDescription": "Brief one-liner",
"category": "development",
"tools": [
{
"name": "search_docs",
"description": "Search documentation by query"
}
],
"envVars": [
{
"name": "API_KEY",
"description": "API key for the service",
"required": true
}
],
"recommendedFor": ["react-expert", "typescript-expert"],
"detectedWhen": ["react", "typescript"]
}
Transport Types
| Transport | When to use | Config |
|-----------|-------------|--------|
| stdio | Local process, development, CLI tools | command + args in .mcp.json |
| http | Cloud services, shared servers | URL endpoint |
| sse | Legacy (deprecated) | URL endpoint |
.mcp.json Configuration
{
"mcpServers": {
"my-server": {
"command": "node",
"args": ["path/to/dist/index.js"],
"env": { "API_KEY": "${API_KEY}" }
}
}
}
Environment variables: ${VAR} syntax, ${VAR:-default} for defaults.
Adding to dev-suite
- Create
mcp-servers/{name}/withpackage.json(@dev-suite/{name}) - Add
metadata.jsonwith tools, envVars, recommendedFor, detectedWhen - Add
src/index.tswith MCP server implementation - Add
tsconfig.json - Update
mcp-servers/package.jsonworkspaces array - Build:
cd mcp-servers && npm install && npm run build
Anti-Patterns
| Anti-Pattern | Fix |
|--------------|-----|
| Huge tool output (> 10K tokens) | Paginate, filter, or summarize results |
| No input validation | Use Zod schemas for all inputs |
| Blocking operations without timeout | Add timeouts to external calls |
| Hardcoded credentials | Use env vars via process.env |
| Tool names that are too generic | Use specific, descriptive names |
Checklist
- [ ] Tools have descriptive names (snake_case) and clear descriptions
- [ ] Input schemas use Zod with validation and defaults
- [ ] Error handling with descriptive messages
- [ ] Environment variables for credentials
- [ ] metadata.json complete (for dev-suite)
- [ ] Package.json added to workspaces
- [ ] Builds successfully:
npm run build - [ ] Output stays under 10K tokens per tool call
Scan to join WeChat group