Skip to content

What Is Tool Use (Function Calling)? A Plain-English Definition

Tool use, or function calling, is how a large language model asks for one of your functions to be run: it returns the function name and its arguments as structured data rather than prose. This is what turns a chatbot into something that can check stock, open a support ticket or save a draft quote. The terms mean the same thing. Anthropic's documentation says "Tool use (also called function calling)", while OpenAI uses function calling and tool calling (as of October 2026).

How does function calling work?

Alongside the user's request, your application sends a list of tools, each with a name, a description and a JSON Schema for its inputs. When a request matches a tool, the model answers with a call instead of text. Your code runs it and returns the result, and the model continues from there.

OpenAI's function calling guide describes five steps:

  1. Send a request that includes the tools the model may call, for example get_order_status with a required order_id.
  2. Receive a tool call with the arguments the model chose.
  3. Run the function in your application, such as a lookup in your order system.
  4. Send a second request that includes the tool output.
  5. Receive the final answer, or further tool calls if the task needs them.

On the Claude API, step 2 shows up as stop_reason: "tool_use" with a tool_use block, and step 4 as a tool_result block. The exception is server tools such as web search, which Anthropic runs on its own infrastructure. Both providers offer a strict mode (strict: true) so that arguments match the schema. Repeat this call-and-result loop several times and you have what is usually called an AI agent.

Tool use vs. MCP vs. API

These three terms sit on different layers. Tool use is the model-side mechanism for requesting an action. An API is the interface of the system that performs it. MCP is a protocol that packages tools once so many AI applications can discover and call them the same way.

  • API: the programming interface of your CRM, ERP or helpdesk. Your code calls it after the model returns a tool call.
  • Model Context Protocol (MCP): an open standard in which a server lists its tools via tools/list and runs them via tools/call (spec version 2026-07-28). The host application turns those tools into function definitions the model can see.
  • Structured outputs: a close relative with no call involved; the model simply answers in a fixed format.

A useful shorthand: tool use decides what to call, MCP standardises how tools are found across applications, and the API does the actual work.

What to watch out for

Because your application executes every call, responsibility for permissions and approvals sits with you, not with the model. Read-only tools carry little risk. Tools that write, send or pay need narrow scopes and, where the stakes are real, a person who confirms.

  • Keep scopes narrow: "save quote as draft" is safer than "post any transaction".
  • Validate inputs: Anthropic notes that some models may guess a value when a required parameter is missing, so check arguments in code.
  • Keep a human in the loop: the MCP specification recommends that a person can always deny a tool invocation.
  • Mind the cost: according to Anthropic, tool names, descriptions and schemas count as input tokens on every request.

For a concrete process view, see how a quotation agent reads master data and prepares drafts without sending anything on its own.

Frequently asked questions

Is tool use the same as function calling?

Yes. Anthropic calls it tool use and names function calling as an alternative term; OpenAI uses function calling and tool calling interchangeably.

Does the model get direct access to my systems?

No. It only produces the call. Your code decides whether to run it, using credentials you control.

Do I need MCP to use function calling?

No. Function calling works directly through the model provider's API. MCP helps when the same tools should be shared across several AI applications.

Are the arguments guaranteed to match my schema?

Only if you enable strict mode (strict: true), which both Anthropic and OpenAI offer. Even then, validate values such as ranges and IDs in your own code.

Sources

  1. Anthropic: Tool use with Claude (accessed 22 Oct 2026) platform.claude.com
  2. OpenAI: Function calling (accessed 22 Oct 2026) developers.openai.com
  3. Model Context Protocol – Specification 2026-07-28: Tools modelcontextprotocol.io