What are tools, MCP and skills, and why does adding more not always help?
Tools, MCP and skills, in plain words.
Open Claude or ChatGPT's settings and you'll find a list of things to switch on: Google Drive, Gmail, Slack, Notion, a calculator, web search.
Each of these is a tool: something the agent can ask the harness to do for it (see "What really happens when an AI agent acts").
The model never sees the app itself. It sees a short description on the desk, like a menu item:
search_drive: Searches your Google Drive. Give it some search words. It returns the names and links of matching files.
That's it. A name, a sentence or two on what it does, and what to give it. The model reads the menu and decides which item to order.
MCP (Model Context Protocol): a standard way to connect a tool to an AI app, so the same tool works in Claude, ChatGPT, Cursor and others.
Think of USB. Before it, every printer and mouse had its own plug. Now one port fits almost anything. MCP is that port for AI tools.
MCP server: a small program that offers one or more tools through that standard plug. "The Slack MCP server" is the thing that lets your AI read and send Slack messages.
Skill: a packet of instructions (and sometimes files) that teaches the AI how to do one kind of task, like "make a slide deck in our company style".
A skill isn't loaded in full all the time. The AI keeps only a one-line summary of each skill on the desk, and opens the full instructions only when a task needs them.
It's like a shelf of manuals with clear spines. You read the spines, and only pull down the one you need.
Every tool you switch on adds its description to the desk, on every single message. Fifty tools is fifty menu items to read before it even starts on your question.
Two things go wrong:
Drew Breunig describes one company that was loading 600 skills every time the agent started up.
A tool's description is text the model reads and trusts. So is a skill.
Add a skill from a stranger and you've put a stranger's instructions on the desk. That's exactly how the Cowork attack in "The lethal trifecta: how AI can leak your data" worked: the hidden instructions came in as a skill.
Switch on only what this job needs. Connect Drive for the Drive task. Turn it off after.
If it keeps picking the wrong tool, remove the overlap. Two search tools for the same place is one too many.
Only add tools and skills from sources you'd trust with the keys. They can read what you can read.