Module 2 of 6 · Advanced · 8 min read
Designing tools and MCP servers
Tools are the agent's hands. Design them for a literal, capable colleague.
- 1Describe when to use a tool, not just what it does.
- 2Return concise outputs and actionable errors.
- 3MCP lets you write a tool server once and use it with any compatible agent.
01Principles of good tools
Tool design is interface design. The model can only use a tool as well as its name, description and schema allow.
- Clear, distinct names: search_orders, not query2.
- Descriptions that say when to use it and when not to.
- Tight schemas with described parameters and sensible defaults.
- Outputs that are complete but concise; paginate long results.
- Errors that explain how to fix the call.
02Fewer, better tools
A dozen overlapping tools confuse models. Consolidate: one well-designed search tool with filters beats five narrow search tools. Build tools around the tasks the agent performs, not around every endpoint in your API.
03The Model Context Protocol
MCP is an open standard, introduced by Anthropic and now supported across OpenAI, Google, Microsoft and most frameworks. An MCP server exposes tools, resources and prompts that any MCP client can use.
Thousands of servers exist for databases, SaaS tools and developer services, so agents can plug into your stack without custom glue.
Finished reading?
Mark it complete to track your progress, then test yourself with the AI Agents quiz.