Your MCP Server is just another client app

An MCP server around your service is another client app for that service. Just as your iOS, Android, or web app are.

This changes everything about how you should build them.

A common mistake

You wouldn't expose raw database operations to your iOS app. You wouldn't make your web app call 15 different microservices to render a dashboard. So why are you making AI agents do that?

Most MCP servers right now are thin wrappers around REST APIs. Every endpoint becomes a tool. Your agent ends up with 40 tools when it needs 5 workflows.

Think like you're building a mobile app

When you built your iOS app, you didn't just add a button per API call. Instead, you:

  • Bundled related operations - One button, multiple API calls behind the scenes
  • Cached and managed state - The app holds context between screens
  • Optimized for the platform - Touch gestures, not mouse hovers

Your MCP server needs the same treatment:

  • Bundle into workflows - get-overview, not get-workspace + get-projects + get-sections
  • Maintain context - Track what the agent's doing across multiple calls
  • Optimize for AI - Design for limited context windows and decision fatigue

What this actually looks like

Here's the Todoist API vs the Todoist MCP server:

API (built for developers):

  • GET /projects
  • GET /sections?project_id=123
  • GET /tasks?section_id=456
  • POST /tasks (one at a time)

MCP Server (built for agents):

  • get-overview - returns entire project structure
  • add-tasks - handles multiple tasks, assigns, schedules

Same backend. Different client needs. Different interface.

The real insight

Your API already has multiple clients. Each one is specialized:

  • Web app: Optimized for browsers, handles sessions, manages routing
  • iOS app: Optimized for touch, handles offline, manages push notifications
  • Android app: Same idea, different platform conventions

Your MCP server is just another client. It should be optimized for how AI agents think and work—not for maximum API flexibility.

Agents need workflows, not endpoints. They need intent-based tools, not CRUD operations.

The Todoist MCP server treats agents as first-class clients: github.com/doist/todoist-ai

If you want to discuss this post, you can do so on Twitter.