tool use
Tool use is the capability by which an AI model calls external programs (search engine, calculator, calendar, database) for help, so it does not just answer from memory but retrieves fresh data, calculates precisely, or performs an operation.
In plain terms
A language model alone is like a very well-read person in an empty room without a phone or internet. They know many things by heart but cannot check today's exchange rates, cannot send an email, and easily make mistakes in complex multiplication. Tool use is when this person is given a phone, calculator, and key to the office.
The tool here is a program the model can invoke: a web search engine, calendar, email client, spreadsheet, code execution environment, or a company's internal system. The model decides whether any are needed for the task, formulates the request, and incorporates the result into its response.
Tool usage is one of the foundations that turns a chatbot into an agent. A simple chatbot only provides text, which you must use. An agent using tools looks things up, calculates, books, and saves on its own, sometimes in multiple sequential steps.
How it works
- The developer provides the model with a list of available tools: each has a name, a short description, and specifies what data it expects (for example, the 'weather' tool expects a city name).
- When you ask something, the model weighs whether it can answer from memory or if a tool is needed.
- If a tool is needed, the model does not write an answer but a well-formed request: which tool to call and with what data. It does not run anything itself; it only asks.
- The surrounding program executes the request (runs the search, queries the database) and returns the result to the model.
- The model reads the result and either writes the answer for you or calls another tool. This loop repeats until the task is completed.
An everyday example
You ask an AI assistant what the weather will be like on Saturday in Balatonfüred and whether it is worth organizing a bike tour. The assistant does not guess but retrieves the forecast from a weather service and, if you have granted permission, checks your calendar, then answers based on this information. On the screen, you see something like: "Requesting weather...", "Checking calendar...".
Who is affected
A common misconception
Limits and risks
- The model may err in which tool to call and with what data, and one bad step can ruin an entire multi-step task.
- A website, email or document read through a tool may contain hidden instructions that hijack the agent (this is prompt injection).
- If the tool can write, send, or pay, an error is no longer just a bad sentence but has real consequences. Therefore, it is advisable to configure the system to request confirmation before irreversible operations.
- Every tool call takes time and money, so a task using many tools is slower and more expensive than a simple response.
Similar, but not the same
- AI-agent – Tool use is a capability, while an agent is the system that independently performs tasks by linking this capability with planning and multiple steps.
- Model Context Protocol (MCP) – MCP is an open protocol for how tools can be connected to models. Tool usage itself is the capability that works even without this protocol.
- plugin – A plugin is a pre-packaged, installable add-on containing tools. Tool use is the mechanism by which the model invokes these tools.
Related concepts
AI-agent, large language model, automation
function calling, web search, code execution, computer use
Model Context Protocol, API, prompt injection, agentciklus, retrieval-augmented generation
A brief history
The concept of a language model alternating between thinking and calling external tools was popularized by the 2022 ReAct study. In early 2023, Toolformer research showed that models can partially learn tool usage themselves. In June 2023, OpenAI embedded the capability into its service under the name "function calling," followed by other major model developers. In November 2024, Anthropic published the Model Context Protocol, which standardizes tool connectivity. In December 2025, Anthropic transferred the protocol to the Agentic AI Foundation, operating under the auspices of the Linux Foundation.
Where things stand
At the end of 2026, virtually every major language model can call tools, and web search or code execution are basic functions in the major chat applications. However, the reliability and security of long tasks involving many tools remain unsolved.
In the news
- Seattle blocks YouTube, Minecraft and Gemini, limits elementary students' screen timeNov 5, 2026 · Education & Family
- Where the human-AI workflow in health care breaks down: hospitals should examine their tech stacks firstNov 5, 2026 · Health
- WorkOS, Northflank and Trigger.dev breached in attacks involving stolen AI access tokensNov 4, 2026 · Code
- Indian authorities warn of AI-powered attacks on financial institutions with unauthorised transfersNov 4, 2026 · Money
- Agent identity: LexisNexis, Mastercard and Trulioo build Know Your Agent toolsNov 4, 2026 · Money
- Google to switch off Search for schools that reject AINov 3, 2026 · Agent Age
- Survey: most parents want AI taught in schools, but with strong safeguardsNov 3, 2026 · Education & Family
- New York City clarifies employer liability for AI-assisted hiring decisionsNov 2, 2026 · Work & HR
Also known as
tool use · tool use · tool calling · tool calling · function calling · function calling · tools · assets
Sources
- Yao et al.: ReAct – Synergizing Reasoning and Acting in Language Models (arXiv, 2022)
- Schick et al.: Toolformer – Language Models Can Teach Themselves to Use Tools (arXiv, 2023)
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.