Tool use

Tool use is when a language model calls an external function or API as part of generating its response. Instead of just producing text, the model can fetch data, run code, query a database, send an email, or trigger any other side effect, by emitting a structured tool-call request that an outside system executes.

This is the foundation of agentic AI. A model with tools is a system that can do things; a model without tools is a system that can only describe doing things.

How it works

Modern LLMs (Claude, GPT-4, Gemini) are trained to:

1. Look at the available tools (provided in the prompt as JSON schemas)
2. Decide whether to call one based on the user’s request
3. Emit a structured request: tool name + arguments
4. Wait for the result
5. Incorporate the result into the next step

Practically, a Claude Code session has access to ~10-20 built-in tools (Read, Edit, Write, Bash, Glob, Grep, etc) plus whatever MCP servers you’ve configured. The model picks tools as it works.

What “tool use” enables for Australian SMB

• File operations. Claude reads your spreadsheets, edits your code, writes new files.
• Shell commands. Claude runs git, npm, your build scripts, custom Bash.
• External APIs. Claude queries Xero, Shopify, your inbox, your database, via MCP.
• Web search + fetch. Claude looks up current information.
• Image generation. Claude (via tool) can call image-gen APIs.

Without tools, the model is a chat window. With tools, it’s a colleague.

Safety considerations

Every tool the model can call is a way for the model to affect the world. The risk profile is the union of all the tool risk profiles.

• Read-only tools (Read, Glob, Grep, web search): low risk, broad usefulness
• Local file edits: medium risk; reviewable via git
• Local shell commands: higher risk; sandbox or restrict
• External writes (sending email, posting to Slack, writing to Shopify): high risk; human approval until trust is established
• Destructive operations (rm, drop database, delete files): very high risk; default deny

Modern AI harnesses let you scope which tools are available per session and per permission level. Use that.

Related terms

• Agent
• Agentic loop
• Model Context Protocol