Tool use and MCP: how agents touch the world
Tools are how an agent gets things done. A tool is a small function the model can call — to search, run code, or talk to other software — and the result comes back for the model to read. The Model Context Protocol (MCP) is a shared open standard that lets any agent plug into any tool server. Think of it as a USB-C port for AI applications.
An agent without tools is just a chatbot. Everything an agent does — reading files, using other software, running code — happens through tools. That makes tool design the least glamorous but most important part of the build.
Here is how tool use works. This is often called function calling — the model asking for a function to be run. The developer lists the functions the model may use. The model writes a request to call one, with the details filled in. The software around the model runs it and hands back the result. The model reads that result and picks its next step. This call-and-check cycle is the loop from guide 01 — tools are its hands.
The part that matters most is the wording. A tool description works like an instruction to the model. The model picks and uses tools based only on their names, descriptions, and notes about their settings. So improving that wording is one of the biggest wins you can get [1]. Anthropic calls this the agent-computer interface. Their advice: a few well-explained tools beat many vague ones. And error messages should tell the model what to try next, not just say something failed [2].
The model only knows what your words tell it — write them with care.
What is MCP, and what does it replace?
Before 2025, every agent needed its own custom connection to every system. The Model Context Protocol (MCP) — an open, shared set of rules for connecting AI apps to outside systems — ended that. A server shares its data, tools, and prompts once, in one format. Then any app that speaks MCP — Claude, ChatGPT, VS Code, Cursor — can connect. The project's own analogy: a USB-C port for AI applications. Build a server once, and every agent can use it [3].
Some practical tips. Give the agent the fewest tools the task needs. Name tools after things the user would recognize. Explain each setting as if writing for a smart colleague who knows nothing about your system. And test your wording the same way you test everything else: run your checks before and after every change.
Sources
- Anthropic — Writing effective tools for agents, 11 Sep 2025
- Anthropic — Building effective agents, 19 Dec 2024 (the agent-computer interface)
- modelcontextprotocol.io — official Model Context Protocol documentation
Frequently asked questions
What is tool use (function calling) in AI agents?
It is how a model takes action. The developer lists the functions the model can use. The model writes a request to call one. The software around the model runs the call and hands back the result. The model reads the result and decides its next step.
What is MCP (Model Context Protocol)?
An open standard — a shared set of rules — for connecting AI apps to outside systems like data sources, tools, and workflows. A server is set up once, and any MCP app (Claude, ChatGPT, VS Code, Cursor) can connect to it. The project calls it a USB-C port for AI applications.
How many tools should an AI agent have?
As few as the task needs. A model can only pay attention to so much. A handful of clearly described tools beats a big pile of vague ones. Every extra tool adds more for the model to read and one more chance to pick wrong. Clear descriptions matter more than the number of tools.