Tool use

Tool use is when an AI model does something instead of only saying something. Rather than producing text and stopping, the model calls an external function - running a shell command, reading or writing a file, querying a database, hitting an API, searching the web - and then reads the result of that call back into its reasoning. This is the capability that turns a language model from a conversation partner into an actor that can change the state of your project.

Tool use is exactly what separates Claude Code from a plain chat window. Claude Code ships with a set of built-in tools - reading files, writing files, running bash commands, searching the codebase, and more - and it reaches for them constantly. Every time it opens one of your files, applies an edit, or runs your test suite, that is tool use happening under the hood. The model is not guessing what your code says; it is using a tool to actually read it, then using another tool to actually change it.

Beyond the built-ins, Claude Code can acquire new tools through MCP servers, which expose external systems - a browser, GitHub, a Postgres database, a search API - over a standard protocol. Each tool the model can call widens what the agent can do without changing the model itself. Tool use also drives the agentic loop: the model uses a tool, observes the result, decides the next step, and repeats. Understanding tool use makes Claude Code's behavior legible - it is not magic, it is a model deciding which function to call next.