Scaffold MCP Server

Scaffold MCP Server

Generate runnable MCP server project from tool spec → official MCP SDK (TS|Py).

Use When

• Have spec (from analyze-codebase-for-mcp or manual) + need server
• New MCP project → right structure from start
• Migrate existing tool integ → MCP protocol
• Prototype tool surface → test w/ Claude Code pre-full-impl
• Need scaffold + test harness for CI

In

• Required: Tool spec doc (YAML|JSON: tool names, params, return types)
• Required: Lang (typescript|python)
• Required: Transport (stdio|sse)
• Optional: Output dir (default: cwd)
• Optional: Pkg name + ver
• Optional: Auth (none|bearer-token|api-key)
• Optional: Docker (true|false, default false)

Do

Step 1: SDK Lang + Transport

1.1. Lang by ctx:

• TS: Node.js, web-adj, JSON-heavy
• Py: Data sci, ML, scientific tools

1.2. Transport:

• stdio: Default local. Claude Code launches as subprocess.
• SSE: Remote|shared. Needs HTTP host.

1.3. Auth:

• none: Local stdio (proc-level trust)
• bearer-token: Remote SSE w/ static tokens
• api-key: Remote w/ per-client keys

→ Lang, transport, auth documented.

If err: ambiguous → default TS + stdio + no auth → fastest time-to-working.

Step 2: Init Project

2.1. Mkdir + init:

TS:

mkdir -p $PROJECT_NAME && cd $PROJECT_NAME
npm init -y
npm install @modelcontextprotocol/sdk zod
npm install -D typescript @types/node tsx
npx tsc --init --target ES2022 --module nodenext --moduleResolution nodenext --outDir dist

Py:

mkdir -p $PROJECT_NAME && cd $PROJECT_NAME
python -m venv .venv
source .venv/bin/activate
pip install mcp pydantic

2.2. Standard structure:

$PROJECT_NAME/
├── src/
│   ├── index.ts|main.py      # Server entry point
│   ├── tools/                 # One file per tool category
│   │   ├── index.ts|__init__.py
│   │   └── [category].ts|.py
│   └── utils/                 # Shared utilities
│       └── validation.ts|.py
├── test/
│   ├── harness.ts|.py         # MCP test harness
│   └── tools/
│       └── [category].test.ts|.py
├── package.json|pyproject.toml
├── tsconfig.json              # TypeScript only
├── Dockerfile                 # If Docker requested
└── README.md

2.3. Bin entry npm (TS) | entry point Py:

TS package.json:

{
  "name": "$PACKAGE_NAME",
  "version": "1.0.0",
  "type": "module",
  "bin": { "$PACKAGE_NAME": "./dist/index.js" },
  "scripts": {
    "build": "tsc",
    "start": "node dist/index.js",
    "dev": "tsx src/index.ts",
    "test": "tsx test/harness.ts"
  }
}

→ Buildable skeleton w/ all deps installed.

If err: install fails → check net + registry. TS → Node ≥18. Py → Py ≥3.10.

Step 3: Tool Handlers from Spec

3.1. Parse spec → per tool, gen handler:

TS template:

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { z } from "zod";

export function registerTools(server: McpServer): void {
  server.tool(
    "tool_name",
    "Tool description from spec",
    {
      param1: z.string().describe("Parameter description"),
      param2: z.number().optional().default(10).describe("Optional param"),
    },
    async ({ param1, param2 }) => {
      try {
        // TODO: Implement tool logic
        const result = await performAction(param1, param2);
        return {
          content: [{ type: "text", text: JSON.stringify(result, null, 2) }],
        };
      } catch (error) {
        return {
          content: [{ type: "text", text: `Error: ${(error as Error).message}` }],
          isError: true,
        };
      }
    }
  );
}

Py template:

from mcp.server import Server
from mcp.types import Tool, TextContent
from pydantic import BaseModel

class ToolNameParams(BaseModel):
    param1: str
    param2: int = 10

async def handle_tool_name(params: ToolNameParams) -> list[TextContent]:
    try:
        result = await perform_action(params.param1, params.param2)
        return [TextContent(type="text", text=json.dumps(result, indent=2))]
    except Exception as e:
        return [TextContent(type="text", text=f"Error: {e}")]

3.2. One handler file per category from spec.

3.3. Validation beyond type checking:

• String len limits
• Numeric bounds
• Enum constraints
• Required field enforce

3.4. Structured err responses for anticipated failures.

→ Handler file per category w/ typed params + err handling.

If err: ambiguous types → default string + TODO for manual refine.

Step 4: Configure Transport

4.1. Server entry w/ chosen transport:

stdio (TS):

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { StdioServerTransport } from "@modelcontextprotocol/sdk/server/stdio.js";
import { registerTools } from "./tools/index.js";

const server = new McpServer({
  name: "$PACKAGE_NAME",
  version: "1.0.0",
});

registerTools(server);

const transport = new StdioServerTransport();
await server.connect(transport);

SSE (TS):

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { SSEServerTransport } from "@modelcontextprotocol/sdk/server/sse.js";
import { registerTools } from "./tools/index.js";

const server = new McpServer({
  name: "$PACKAGE_NAME",
  version: "1.0.0",
});

registerTools(server);

const transport = new SSEServerTransport("/messages", response);
await server.connect(transport);

4.2. Auth req → middleware:

• Bearer token → validate Authorization header
• API key → validate X-API-Key header

4.3. Shebang for stdio → direct exec:

#!/usr/bin/env node

→ Working entry starts MCP server on transport.

If err: SDK ver ≠ import paths → check @modelcontextprotocol/sdk ver + adjust. SDK restructured paths between vers.

Step 5: Test Harness

5.1. Validate every tool:

import { McpServer } from "@modelcontextprotocol/sdk/server/mcp.js";
import { InMemoryTransport } from "@modelcontextprotocol/sdk/inMemory.js";
import { Client } from "@modelcontextprotocol/sdk/client/index.js";

async function runTests(): Promise<void> {
  const server = createServer();
  const [clientTransport, serverTransport] = InMemoryTransport.createLinkedPair();

  await server.connect(serverTransport);
  const client = new Client({ name: "test-client", version: "1.0.0" });
  await client.connect(clientTransport);

  // Test: tools/list returns all expected tools
  const tools = await client.listTools();
  console.assert(tools.tools.length === EXPECTED_TOOL_COUNT);

  // Test: each tool with valid input
  for (const tool of tools.tools) {
    const result = await client.callTool({
      name: tool.name,
      arguments: getTestInput(tool.name),
    });
    console.assert(!result.isError, `${tool.name} failed`);
  }

  // Test: each tool with invalid input returns isError
  for (const tool of tools.tools) {
    const result = await client.callTool({
      name: tool.name,
      arguments: getInvalidInput(tool.name),
    });
    console.assert(result.isError, `${tool.name} should reject invalid input`);
  }

  console.log("All tests passed");
}

5.2. Test fixtures per tool: valid, invalid, edge cases.

5.3. Add test script → package.json | pyproject.toml.

→ Harness exercises every tool w/ valid+invalid.

If err: InMemoryTransport not in SDK ver → fall back to spawning server as subproc + stdio pipes.

Step 6: Docs + Config

6.1. Gen README.md w/:

• Project desc
• Install instructions
• Claude Code config cmd
• Claude Desktop JSON snippet
• Tool listing w/ descs + param schemas
• Dev + test instructions

6.2. Gen Claude Code register cmd:

# stdio transport
claude mcp add $PACKAGE_NAME stdio "node" "dist/index.js"

# SSE transport
claude mcp add $PACKAGE_NAME -e API_KEY=your_key -- mcp-remote http://localhost:3000/mcp

6.3. Gen Claude Desktop config:

{
  "mcpServers": {
    "$PACKAGE_NAME": {
      "command": "node",
      "args": ["path/to/dist/index.js"]
    }
  }
}

6.4. Docker requested → gen Dockerfile:

FROM node:20-slim AS build
WORKDIR /app
COPY package*.json ./
RUN npm ci
COPY . .
RUN npm run build

FROM node:20-slim
WORKDIR /app
COPY --from=build /app/dist ./dist
COPY --from=build /app/node_modules ./node_modules
COPY --from=build /app/package.json .
ENTRYPOINT ["node", "dist/index.js"]

→ Complete docs + config for immediate use.

If err: README has placeholders → search project for actual vals to substitute. Docker fail → verify base img matches Node|Py ver.

Check

• Builds w/o errs (npm run build|equiv)
• Server starts + responds tools/list JSON-RPC
• Every tool from spec registered + discoverable
• Harness passes valid inputs all tools
• Harness confirms err responses invalid inputs
• Claude Code connects via claude mcp add
• README has working install + config
• Gen code passes lint (if configured)

Traps

• SDK import path changes: @modelcontextprotocol/sdk restructured between vers. Check installed ver export paths.
• Forget shebang: stdio invoked directly needs #!/usr/bin/env node first line.
• Block event loop: TS handlers must async. Sync ops block all other tool calls.
• Missing type: "module": SDK uses ESM. Without → Node treats as CJS, imports fail.
• Zod schema drift: Spec evolves but Zod not updated → silent fails. Gen schemas from single source of truth.
• stdout pollution: stdio uses stdout for JSON-RPC. Any console.log in handlers corrupts. Use console.error|file logger.

→

• analyze-codebase-for-mcp — gen spec this skill consumes
• build-custom-mcp-server — manual impl for complex
• configure-mcp-server — connect scaffold → Claude Code/Desktop
• troubleshoot-mcp-connection — debug post-deploy
• containerize-mcp-server — Docker for distribution