Tool Use
Also known as: Tool Usage, Tool Calling, External Tools
Definition
The capability of an LLM to invoke external functions, APIs, or services to perform actions or retrieve information beyond text generation. Tool use extends LLMs from knowledge systems to action systems—they can search the web, run code, query databases, send emails, and interact with any API. The model decides when to use which tool and formulates appropriate inputs.
What this is NOT
- Not the tools themselves (tool use is the act of invoking tools)
- Not RAG (RAG is retrieval; tool use includes actions beyond retrieval)
- Not all structured output (tool use implies execution, not just formatting)
Alternative Interpretations
Different communities use this term differently:
llm-practitioners
A model capability where the LLM outputs structured tool invocations (often JSON) that are executed by the application, with results fed back to the model. Enabled via function calling APIs or tool-use prompts.
Sources: OpenAI Function Calling documentation, Anthropic Tool Use documentation, Google Gemini function calling
cognitive-science
The use of external artifacts to extend cognitive or physical capabilities. Tool use is a marker of intelligence in animals and is fundamental to human cognition (extended mind thesis).
Sources: Clark & Chalmers: The Extended Mind, Tool use in animal cognition research
Examples
- Model calls a weather API to answer 'What's the weather in Paris?'
- Model executes Python code to calculate statistics
- Model searches the web to find recent information
- Model sends an email on behalf of the user
Counterexamples
Things that might seem like Tool Use but are not:
- Model generating a response from its training knowledge (no tool invoked)
- RAG retrieval (retrieval is passive; tool use implies active invocation)
- Model outputting JSON without execution (that's structured output)
Relations
- requires function-calling (Function calling is the API mechanism for tool use)
- requires tool-definition (Tools must be defined for the model)
- produces tool-result (Tool invocations produce results)
- overlapsWith agent (Agents use tools to take actions)
Implementations
Tools and frameworks that implement this concept:
- Anthropic primary
- Claude Chrome Extension overlapping
- Context7 secondary
- Ghostty overlapping
- Playwright primary
- Playwright MCP Server primary
- Superpowers (obra/superpowers) secondary
- Visual Studio Code secondary