MCP Servers: Prompting Agents With the Right Context

What MCP Adds

The Model Context Protocol (MCP) lets agents call tools—databases, GitHub, browsers, custom APIs—with structured permissions. Prompting must say when to use tools and how to cite outputs.

Tool Discipline Prompt

"You have MCP tools for Postgres and GitHub. Rules: (1) Query DB only for read-only diagnostics unless I say WRITE. (2) Cite issue numbers when referencing tickets. (3) If a tool errors, show exact error text and stop—do not fabricate data."

Investigation Workflow

"Step 1: search GitHub issues for 'checkout timeout'. Step 2: run read-only SQL to count failed payments last 24h. Step 3: correlate findings. Output table: Source | Finding | Confidence."

Least Privilege Reminder

Tell the agent which tools are allowed this session. Forbid destructive Git or production writes by default.

Human Handoff

"End with: decisions needed from human, suggested next commands, and data freshness (timestamps)."