Tool use overcomes a lot of the core limitations of LLMs...Maybe you give it a calculator so it doesn’t have to rely on its internal, unreliable arithmetic abilities. Maybe you let it search the web or view your calendar, or you give it (read-only!) access to a company database so it can pull up information or search technical documentation.
A language model using external tools as part of a reasoning loop, via Tools Fail: Detecting Silent Errors in Faulty Tools (Sun et al.)
Tool use in LLMs refers to a model's ability to access and invoke external software — such as APIs, databases, search engines, or other functions — in order to accomplish a task.
Language models are stateless text predictors. But with tool use, they can:
- Run searches
- Call APIs
- Access live data
- Perform calculations
- Control applications or workflows
This turns them into agents: systems that reason, plan, and take action using a growing toolkit.
How it works:
- Function signatures are defined and passed into the model
- The model predicts which tool to call (and with what arguments)
- The developer or runtime executes the tool
- The response is optionally returned to the model for further reasoning
Most tools are exposed via structured formats like JSON, and protocols like JSON-RPC or MCP enable secure, model-agnostic integrations.
Why this matters:
Tool use lets LLMs solve real problems that require up-to-date or external knowledge — making them interactive, composable, and extensible.
FAQ
What is tool use in LLMs?
How do LLMs decide which tool to use?
Is LLM tool use secure?
What kinds of tools can be used by LLMs?
Related Stuff
- Function Calling: How models choose and call external functions using structured JSON
- Model Context Protocol: An open standard for model-to-tool communication
- Retrieval Augmented Generation: Tool use example: LLMs querying external documents for better answers