Start a project
Practice Reviewed 24 Sep 2026 · 17 sources

tool use

/ˈtuːl ˈjuːs/ noun Verb use tools · noun tool call · adjective tool-using
Definition

An AI model’s ability to request that the software around it run a described function, such as a web search or a calculation, and then use the result.

Examples
  • A chatbot asking a weather service for Boston’s current conditions
  • A model running Python to work out compound interest
  • A sales assistant querying a database for last month’s orders
  • An email to a colleague, sent only after the user confirms
A model hands a filled-in request slip reading weather, Boston to an application. The application runs it against a weather service and passes the result back to the model.
Inkling · The model’s request, carried out by the applicationFig. T1
Earliest AI use
2005
Earliest found: an AAAI paper on autonomous agents that use physical tools.
Current meaning
2022
Google’s LaMDA checks its claims with a search system and a calculator.
Also called
Function calling
OpenAI’s name, from June 2023. Its guide also says tool calling.
§01

In plain words

Tool use, or function calling, is an AI model’s ability to have outside software, such as a web search or calculator, run for it and return a result.

On its own, a large language model can only produce text from its training and its prompt. It cannot check today’s weather or look up a customer’s order. It is also weak at arithmetic and factual lookup, which much simpler programs do well.1

Tool use connects the model to such programs. The developer describes each tool: its name, what it does and the inputs it takes. When a request needs one, the model replies with a structured call instead of an answer. The application runs the call and returns the result, and the model writes its reply from it.2 Asked what $100 becomes at 7% compound interest over 12 years, a model with a code tool can compute it rather than guess.3

OpenAI launched the capability as function calling in June 2023, and its guide also says tool calling. Anthropic calls it tool use.

Tool use is the basic step inside an AI agent. Anthropic describes agents as models using tools in a loop, each result shaping the next step.4 Tool use is also confused with the Model Context Protocol (MCP), a shared standard for offering functions from separate servers. A model can use tools without MCP.

Tools that act in the world add risk. When OpenAI launched function calling, it warned that untrusted content in a tool’s output could steer the model into unintended actions. That is a form of prompt injection. OpenAI advised asking users to confirm actions such as sending an email, which keeps a human in the loop.5

§02

How it works

A tool call takes at least two requests to the model.2 In the first, the application sends the user’s message and the list of tools. Each tool has a name, a description and a schema: a formal list of the inputs it accepts, written in JSON Schema. JSON is a plain-text format for structured data.

If a tool would help, the model stops and returns a tool call: the tool’s name and the input values, in JSON. The application runs the function, such as a web request or a database query. In the second request it sends back the output, and the model answers or calls another tool.

Server tools work differently. Anthropic, for example, runs tools such as web search and code execution on its own systems and returns the result directly.6

Deciding when to call a tool is a learned skill. In 2023 Toolformer showed that a model could teach itself which tool to call, when to call it and with what inputs.1 ReAct, in 2022, had a model alternate between reasoning and actions, reading each result before the next step. Agentic systems repeat that loop over many steps.

§03

Origin and first known use

Tool use is older as a term in animal behavior, where it describes an animal handling an object to reach a goal. Jane Goodall reported tool-using chimpanzees in Nature in 1964. AI researchers applied the phrase to robots and software agents; a 2005 AAAI paper found it little studied in AI. In 2019 OpenAI reported emergent tool use by agents playing hide-and-seek. The software sense arrived with language models: Google’s LaMDA consulted a toolset in January 2022. OpenAI named its API feature function calling in June 2023, and Anthropic launched its own as tool use in 2024.

  1. 1964 Jane Goodall reports tool-using wild chimpanzees in Nature. The phrase’s older sense, from animal behavior, predates its use in AI.
    Goodall, “Tool-Using and Aimed Throwing in a Community of Free-Living Chimpanzees”, Nature 201, 28 March 1964 nature.com →: source for 1964
  2. 2005 St. Amant and Wood’s AAAI paper “Tool Use for Autonomous Agents” finds tool use little studied in AI and draws on animal cognition. Earliest AI use found.
    St. Amant and Wood, “Tool Use for Autonomous Agents”, AAAI 2005 aaai.org →: source for 2005
  3. 2019 OpenAI researchers report hide-and-seek agents that learn to build shelters from movable boxes and get past obstacles with ramps, and call it emergent tool use.
    Baker et al., “Emergent Tool Use From Multi-Agent Autocurricula” arXiv 1909.07528 →: source for 2019
  4. 2022 Google’s LaMDA paper gives a dialog model a toolset of an information retrieval system, a calculator and a translator, for looking up its claims.
    Thoppilan et al., “LaMDA: Language Models for Dialog Applications” arXiv 2201.08239 →: source for 2022
  5. 2022 ReAct has a language model interleave reasoning with actions, such as calls to a simple Wikipedia API, and read each result before its next step.
    Yao et al., “ReAct: Synergizing Reasoning and Acting in Language Models” arXiv 2210.03629 →: source for 2022
  6. 2023 Toolformer teaches itself when to call a calculator, a Q&A system, two search engines, a translator and a calendar, from a few demonstrations each.
    Schick et al., “Toolformer: Language Models Can Teach Themselves to Use Tools” arXiv 2302.04761 →: source for 2023
  7. 2023 OpenAI launches function calling in its API. Developers describe functions, and the model can return a JSON object with the arguments to call them.
    OpenAI, “Function calling and other API updates”, 13 Jun 2023 openai.com →: source for 2023
  8. 2024 Anthropic makes tool use generally available for its Claude 3 models, on its own API, Amazon Bedrock and Google Cloud’s Vertex AI.
    Anthropic, “Claude can now use tools”, 30 May 2024 claude.com →: source for 2024
§04

In the wild

8 sightings · 2019–2026 · oldest first
  1. Paper17 Sep 2019

    “…we find that agents create a self-supervised autocurriculum inducing multiple distinct rounds of emergent strategy, many of which require sophisticated tool use and coordination.”

    Baker et al., arXiv arXiv 1909.07528 →: Baker et al., arXiv, 17 Sep 2019Archived →: copy of Baker et al., arXiv, 17 Sep 2019
  2. Headline17 Sep 2019

    “AI learned to use tools after nearly 500 million games of hide and seek”

    MIT Technology Review technologyreview.com →: MIT Technology Review, 17 Sep 2019Archived →: copy of MIT Technology Review, 17 Sep 2019
  3. Blog13 Jun 2023

    “Function calling allows developers to more reliably get structured data back from the model.”

    OpenAI, product announcement openai.com →: OpenAI, product announcement, 13 Jun 2023Archived →: copy of OpenAI, product announcement, 13 Jun 2023
  4. Blog23 Jun 2023

    “Tool use is a remarkable and distinguishing characteristic of human beings.”

    Lilian Weng, on her blog Lil’Log lilianweng.github.io →: Lilian Weng, on her blog Lil’Log, 23 Jun 2023Archived →: copy of Lilian Weng, on her blog Lil’Log, 23 Jun 2023
  5. Paper25 Sep 2023

    “…Language Model (LM) agents and tool use, exemplified by applications like ChatGPT Plugins, enable a rich set of capabilities but also amplify potential risks…”

    Ruan et al., arXiv arXiv 2309.15817 →: Ruan et al., arXiv, 25 Sep 2023Archived →: copy of Ruan et al., arXiv, 25 Sep 2023
  6. Blog3 Apr 2024

    “Tool Use, in which an LLM is given functions it can request to call … is a key design pattern of AI agentic workflows.”

    Andrew Ng, in The Batch deeplearning.ai →: Andrew Ng, in The Batch, 3 Apr 2024Archived →: copy of Andrew Ng, in The Batch, 3 Apr 2024
  7. Blog30 May 2024

    “Tool use, which enables Claude to interact with external tools and APIs, is now generally available across the entire Claude 3 model family…”

    Anthropic, product announcement claude.com →: Anthropic, product announcement, 30 May 2024Archived →: copy of Anthropic, product announcement, 30 May 2024
  8. Docs25 Aug 2026

    “The model generates function calls and structured arguments based on the functions you specify.”

    Microsoft, Foundry Models documentation learn.microsoft.com →: Microsoft, Foundry Models documentation, 25 Aug 2026Archived →: copy of Microsoft, Foundry Models documentation, 25 Aug 2026
§05

Not to be confused with

  • Model Context Protocol (MCP)

    Tool use is a model’s ability to ask for a described function to run. MCP is an open standard for offering such functions from separate servers to any compatible application. A model can use tools without MCP.

    See Model Context Protocol →

  • AI agent

    An AI agent pursues a goal over many steps, typically calling tools in a loop. Tool use is one capability: requesting a function. A chatbot that checks the weather once uses a tool without being an agent.

    See AI agent →

  • ChatGPT plugins

    Plugins, launched by OpenAI in March 2023, were tools packaged for one product, ChatGPT, each described in a manifest and an OpenAPI file. Tool use is the general ability of a model to call tools.

  • Tool use (animal behavior)

    In biology, tool use is an animal handling an object, such as a stick, to reach a goal like food. The AI sense borrows the phrase for software a model can ask to run.

§06

Questions

  • What is the difference between tool use and function calling?

    There is no difference in meaning. They are two names for the same capability. OpenAI launched it as function calling in June 2023, and its guide also calls it tool calling. Anthropic calls it tool use. In each case the model asks for a named function to run with specific inputs. The software around the model runs it and returns the result.

  • Does the model run the tool itself?

    Usually not. The model chooses a tool and writes its inputs as a structured call. The application that sent the request runs the function, such as a database query, and returns the result for the model to use. Server tools are the exception. Anthropic, for example, runs tools such as web search and code execution on its own systems and returns the result directly.

  • Is tool use the same as an AI agent?

    No. Tool use is one capability: asking for a function to run and reading the result. An AI agent is a system that works toward a goal over many steps. Anthropic describes agents as models using tools in a loop, guided by feedback from each result. A chatbot that checks the weather once uses a tool without being an agent.

  • What is the difference between tool use and MCP?

    Tool use is a model’s ability to ask for a described function to run. The Model Context Protocol (MCP) is a shared standard for offering such functions from separate servers. One set of tools can then serve many applications. An application can also define its tools itself, with no server involved. A model can therefore use tools without MCP.

  • Is tool use safe?

    It carries risks, especially when tools can take actions. Text returned by a tool can contain instructions that steer the model into unintended actions, a form of prompt injection. OpenAI warned of this when it launched function calling in June 2023. It advised asking users to confirm actions such as sending an email. A 2023 research paper also argued that tool use amplifies potential risks.

§07

Watch

§09

Cite this entry

“Tool use.” SUPERR_dictionary, Superrduperr, 24 Sept. 2026, superrduperr.com/dictionary/tool-use.

First published 24 Sep 2026 · Last reviewed 24 Sep 2026 · Report an error →