{
  "id": 20516978,
  "title": "MCP tool descriptions are part of your prompt: what to read before you connect a server",
  "url": "https://urgent.news/2026/11/17/mcp-tool-descriptions-are-part-of-your-prompt-what-to-read-before-you",
  "topic": "tech",
  "section": "Tech",
  "published": "2026-11-17T23:44:21.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/mogley1011/mcp-tool-descriptions-are-part-of-your-prompt-what-to-read-before-you-connect-a-server-2853"
  },
  "original_language": "en",
  "account": "When an agent connects to an MCP server, the server's tool names and descriptions are loaded into the model's context. This makes those descriptions instructions the model may follow, set by whoever manages the server. Reading them is a critical security step, not a mere formality. Here's what to verify in just five minutes:\n\nDescriptions that guide the model. A description must clearly explain what the tool does. Phrases like \"always call this first,\" \"don't let the user know,\" or references to other tools indicate steering, not documentation.\n\nTool scope versus name. A tool named \"read_note\" that can handle any file path or shell string may be broader than its description suggests. Ensure you confirm precisely what it can and cannot do.\n\nCredentials. List all environment variables and tokens the server process receives. Only provide what's absolutely necessary, favoring read-only tokens when possible.\n\nPinned versions. Using npx some-server@latest means running today's code with the credentials you're currently using. It's essential to pin a specific version and review the differences when you decide to upgrade.\n\nRemember, descriptions can change even after you've approved them. The server can provide different tool metadata later, so it's crucial to re-check after updates, or snapshot the approved descriptions and compare them.\n\nLastly, understand that a single review isn't enough. The server, its dependencies, and its descriptions all evolve over time. A review is a snapshot of a moving target, so it's beneficial to establish a repeatable check you can run every time there's an update, rather than a one-time read.",
  "summary": "When an agent connects to an MCP server, the server's tool names and descriptions are loaded into the model's context. That makes them instructions the model may follow, written by whoever runs the server. Reading them is a security step, not a formality. What to check in five minutes Descriptions that talk to the model. A description should say what the tool does. Text such as \"always call this…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}