← ClaudeAtlas

mcp-serverlisted

Design a Model Context Protocol server that exposes tools and data to AI agents safely and legibly. Use when building an MCP server or deciding what to expose to an agent.
Amey-Thakur/AI-SKILLS · ★ 4 · AI & Automation · score 77
Install: claude install-skill Amey-Thakur/AI-SKILLS
# MCP server The Model Context Protocol lets an agent discover and call your tools and read your resources. The server is a contract with a caller that reasons in natural language and will misuse anything ambiguous, so design for a smart caller that cannot see your source. ## Method 1. **Expose the right primitive.** MCP offers tools (actions the agent invokes), resources (data the agent reads), and prompts (templates the agent fills). Read-only data is a resource, not a tool; an action with effects is a tool. Choosing wrong makes the agent fetch data by calling side-effecting tools, or take actions it thinks are safe reads. 2. **Name and describe for the model, not the maintainer.** Every tool's name and description is how the agent decides whether and how to call it. Say what it does, when to use it, and what it returns, in one clear statement. A vague description produces wrong calls no schema can prevent. 3. **Make schemas strict and self-documenting.** Each parameter has a type, a description, and required-or-optional stated. Constrain enums, ranges, and formats so the model cannot pass a shape you will reject. Invalid states the schema forbids are calls you never have to handle. 4. **Return results the model can use.** Structured, labeled output over raw dumps; the fields the agent needs to decide the next step, not your internal representation. Truncate or paginate large results so one call cannot blow the context window. 5. **