MCP API Build

MCP API Build

Design and build production APIs and MCP servers by corporate standards.

When to Use

• Designing a new REST API from scratch
• Wrapping an existing REST API with MCP server
• Converting REST endpoints → MCP tools/resources/prompts
• Writing OpenAPI spec (API-first approach)
• Reviewing/auditing API architecture and anti-patterns
• Planning API security and production deployment

Workflow

Path A: Build REST API from scratch

1. Domain Model → 2. Resource Inventory → 3. OpenAPI Spec → 4. Implement → 5. Audit → 6. Deploy

1. Domain Model: Identify entities, relationships, consumers. Ask: who calls API? what problems does it solve? what tiers/permissions exist?
2. Resource Inventory: List resources + CRUD matrix. See rest-api-standards.md §Design Process Checklist
3. OpenAPI Spec: Write spec BEFORE code. See openapi-first-process.md
4. Implement: Code against contract. Any deviation = bug
5. Audit: Check anti-patterns. See advanced-patterns.md §Anti-Patterns
6. Deploy: Security + production checklist. See security-production.md

Path B: Build MCP server for existing API

1. Audit REST API → 2. Map endpoints → 3. Write server → 4. Add advanced features → 5. Test → 6. Deploy

1. Audit existing API: List all endpoints, methods, params, responses
2. Map to MCP primitives: Use decision matrix in rest-to-mcp-mapping.md
3. Write MCP server: Follow patterns in mcp-server-patterns.md
4. Advanced features: Progress notifications, elicitation, structured output. See advanced-patterns.md §Advanced MCP Patterns
5. Test: MCP Inspector + programmatic tests
6. Deploy: stdio (local) or Streamable HTTP (remote)

Path C: Design both REST API + MCP together

Combined path. Design REST first (Path A steps 1-3), then layer MCP on top (Path B steps 2-6).

Quick Reference: REST Design Rules

• Plural nouns: /users, /events, /tickets
• HTTP methods = actions: GET read, POST create, PATCH update, DELETE remove
• No verbs in URL: POST /tickets not /createTicket
• Max 2 nesting levels: /users/{id}/orders ok, deeper = flatten
• snake_case JSON fields: created_at, user_id
• Consistent responses: same shape for all lists, all errors
• Paginate everything: limit + offset or cursor
• Version in URL: /v1/
• Idempotency keys for POST: Idempotency-Key: req_abc123
• ETag for caching/concurrency: If-Match, If-None-Match
• Health check: GET /health always

Full reference: rest-api-standards.md

Quick Reference: MCP Primitives

Decision: GET without params → Resource. GET with params or any write → Tool.

Full spec: mcp-specification.md

Quick Reference: MCP Tool Design

Name:        snake_case verb_noun (search_knowledge, create_ticket)
Description: What it does + what it returns + when to use (for LLM)
inputSchema: JSON Schema with description on every field, enums where possible
Output:      Structured JSON (structuredContent), not stringified text
Errors:      Return isError: true, not exceptions
Annotations: readOnlyHint, destructiveHint, idempotentHint

Max 10-15 tools per server. Group logically. Write comprehensive server instructions.

Full mapping guide: rest-to-mcp-mapping.md

Quick Reference: Advanced Features (2025-06-18+)

References (load on demand)

Example: Production MCP Server (Python)

import os
import httpx
from mcp.server.fastmcp import FastMCP

BASE_URL = os.getenv("API_BASE_URL", "https://api.myapp.com/v1")
API_KEY = os.getenv("API_KEY")
HEADERS = {"Authorization": f"Bearer {API_KEY}"}

mcp = FastMCP("myapp", instructions="""
MyApp MCP server -- educational platform for AI practitioners.

## Tools
- search_knowledge: Find learning materials (lessons, skills, workshops)
- create_ticket: Report issues or request help
- get_event: Get event details with calendar link

## Guidelines
1. Search before listing (faster)
2. Check material tier before recommending
3. For problems, create a ticket
""")

async def _api(method: str, path: str, **kwargs) -> dict:
    """Internal API caller with timeout and error handling."""
    async with httpx.AsyncClient(base_url=BASE_URL, timeout=30.0) as c:
        r = await getattr(c, method)(path, headers=HEADERS, **kwargs)
        if r.status_code >= 400:
            return {"error": r.json().get("error", {"code": "unknown"})}
        return r.json()

@mcp.tool(annotations={"readOnlyHint": True, "openWorldHint": True})
async def search_knowledge(
    query: str,
    category: str | None = None,
    limit: int = 20
) -> dict:
    """Search knowledge base by keywords.
    Returns lessons, skills, usecases, workshops, guides.
    Use when user asks about learning materials or tutorials."""
    params = {"q": query, "limit": min(limit, 100)}
    if category:
        params["category"] = category
    return await _api("get", "/knowledge/search", params=params)

@mcp.tool(annotations={"destructiveHint": False, "idempotentHint": False})
async def create_ticket(subject: str, body: str) -> dict:
    """Create support ticket. Use when user reports a problem."""
    return await _api("post", "/support/tickets",
                      json={"subject": subject, "body": body})

@mcp.resource("myapp://profile/me")
async def get_profile() -> dict:
    """Current user profile and subscription tier."""
    return await _api("get", "/profile/me")

@mcp.resource("myapp://notifications/unread")
async def unread_count() -> dict:
    """Number of unread notifications."""
    return await _api("get", "/notifications/unread-count")

Run: python server.py | Test: npx @modelcontextprotocol/inspector python server.py