MCP servers range from thin wrappers to purpose-built tool sets. Choosing the right pattern affects how well models can use them.
Thin API Wrapper
Each API endpoint becomes a tool. Quick to build, but models may need many calls and must understand the API's quirks.
Task-Oriented Tools
Tools shaped around what users want to accomplish: summarise_customer_account rather than five separate lookups. Fewer calls, better results, but more design work.
Read-Only Data Access
Expose search and read operations only. Lower risk; a good starting point for sensitive systems.
Gateway or Aggregator
One server fronting several back-end systems, with shared authentication, logging and policy enforcement. Useful for organisations managing many integrations.
Code Execution
Rather than many individual tools, let the model write code against an API in a sandbox. Powerful and flexible, but needs strong isolation.
Choosing
- Start read-only and task-oriented.
- Add write tools deliberately, with clear descriptions and approval.
- Keep the number of tools manageable; very large tool lists confuse models and consume context.
Versioning
Change tools carefully; hosts and prompts may depend on names and schemas.