As AI agents become first-class users of our platforms, many teams are discovering something surprising:
Our traditional REST APIs are not always the best interface for them.
REST was designed for browsers, servers, and mobile apps.
AI agents — especially those built on protocols like Model Context Protocol (MCP) — interact with systems very differently.
AI agents — especially those built on protocols like Model Context Protocol (MCP) — interact with systems very differently.
This shift has created a new pattern:
Agent-specific tools and resources, which run alongside traditional REST endpoints, not instead of them.
Agent-specific tools and resources, which run alongside traditional REST endpoints, not instead of them.
1. Traditional REST Resources: Built for Apps and Humans
REST has been the backbone of APIs for years. It’s predictable, stable, cache-friendly, and great for humans to reason about.
Example REST endpoint:
GET /api/v1/products?query=cement
Strengths of REST
- Standardised patterns (CRUD, pagination, filters)
- Well-understood by developers, tools, SDKs
- Easy to secure (tokens, sessions, headers)
- Easy to test and integrate with frontend apps
- Data-focused — returns structured JSON
REST is perfect when you're building:
- Frontend / mobile clients
- Integrations for partners
- Internal dashboards
- Reporting tools
But REST starts to show limitations when the consumer is not a human developer — but an AI agent.
2. Agent-Specific Tools & Resources: Built for ML Models
AI agents don’t think in terms of CRUD APIs.
They think in terms of:
- Tasks
- Operations
- Goals
- Plans
For example, an agent doesn’t want a “GET endpoint”.
It wants a tool like:
It wants a tool like:
product_search(query: string) → list of products
Agents need:
Higher-level abstractions
A tool that encapsulates business logic, not just raw data.
Safer, constrained operations
Agents should not be able to PUT, PATCH or delete arbitrary resources.
Natural-language-aligned actions
Tools map to intents, not endpoints.
Tools map to intents, not endpoints.
Well-structured, predictable outputs
Agents need stable, verifiable data shapes — not every field you might return to a frontend.
Agents need stable, verifiable data shapes — not every field you might return to a frontend.
Stateless or lightly-stateful interactions
Tools map to requests, not resource state transitions.
This is where MCP tools come in.
Tools map to requests, not resource state transitions.
This is where MCP tools come in.
3. The Core Difference: “Resources” vs “Capabilities”
REST: Resources
REST exposes things:
- products
- orders
- customers
- quotes
These are nouns, entities, and collections.
MCP Tools: Capabilities
MCP exposes actions:
- product_search
- estimate_delivery_window
- compare_quotes
- check_stock_level
These are verbs — tasks and operations.
Agents understand verbs far better than nouns.
4. Why REST Alone Isn’t Enough for Agents
Let’s say you're exposing product search to an agent.
REST version:
GET /api/v1/products?query=cement&category=bagged
REST returns every field, pagination metadata, pricing fields that don’t matter, etc.
AI agents struggle with:
- overly complex payloads
- noisy fields
- inconsistent shapes
- nested objects
- needing multiple requests to accomplish one job
MCP tool version:
tools/product_search:
input: { query: string }
output: { products: [{ name, price, sku, pack_size }] }
The agent gets only what it needs to perform a task.
Not a database dump.
Not a database dump.
5. Why Keeping Both Is Powerful
You don't want to replace REST.
REST powers your customers, your frontend, your partners, and your long-lived ecosystem integrations.
REST powers your customers, your frontend, your partners, and your long-lived ecosystem integrations.
But agents benefit from a different interface that is:
- more intentional
- safer
- more constrained
- easier for LLMs to interpret
- aligned with natural language workflows
Analogy
- REST is like exposing a database over HTTP.
- MCP tools are like exposing a concierge that performs tasks for you.
Both coexist beautifully.
6. The Big Picture: REST + MCP = Future-Proof APIs
By keeping REST resources and agent tools separate, you get the best of both worlds:
REST
For humans and systems that want stable, CRUD-like, predictable resource access.
For humans and systems that want stable, CRUD-like, predictable resource access.
MCP Tools
For AI agents that want to perform tasks safely, consistently, and efficiently.
This dual-layer API design is becoming a standard pattern in systems adopting LLMs.
For AI agents that want to perform tasks safely, consistently, and efficiently.
This dual-layer API design is becoming a standard pattern in systems adopting LLMs.
Conclusion
REST APIs aren’t going anywhere — they’re foundational.
But AI agents require a different interface, one that reflects actions instead of resources.
But AI agents require a different interface, one that reflects actions instead of resources.
By introducing MCP tools alongside traditional REST resources, you give:
- Apps: a clean API
- Agents: a safe, structured, task-based interface
- Your platform: flexibility to evolve
The future of APIs isn’t REST vs MCP —
it’s REST and MCP, working together.
it’s REST and MCP, working together.