M
MCP Directory CURATED DIRECTORY
693 MCP servers indexed
中文

Function / Tool Calling

Function / Tool Calling

Also called: 函数调用 · 工具调用 · Tool use

The model's ability to request an external function in a structured format and continue once it has the result — the basis of nearly all agent capability.

Model emitsJSON callRuntimeexecutes fnResult backto modelstructured, schema-checked
Schematic (simplified)

The flow of one call

  1. The app includes a tool manifest in the request (each tool's name, purpose, and parameter JSON Schema).
  2. The model does not run code; it emits structured data: "call get_weather with {city: "Paris"}."
  3. Application code actually executes that function and feeds the return value back as a new message.
  4. The model continues from the result — possibly calling another tool, or giving a final answer.

Why it is pivotal

It turns the model from "generates text only" into "can trigger deterministic operations and read real results." Agent loops, MCP, and retrieval augmentation are all built on this mechanism. Constraining arguments with a schema makes calls checkable, reliable, and reproducible.

Common misconceptions

  • The model does not execute the function — it produces a "call intent"; execution is your runtime's job.
  • Whether a tool call succeeds depends heavily on how well the tool is described — vague descriptions lead to wrong tools or wrong arguments.

Related terms

Tool Manifest Tool ManifestMCP (Model Context Protocol) MCP (Model Context Protocol)Agent Loop Agent LoopAI Agent AI Agent

Sources

  1. Anthropic — Claude Docs: Tool use (function calling)
  2. OpenAI — Function calling guide

Compiled 2026-08-29 · This glossary is compiled from public papers, official specifications, and common industry definitions, and is updated as the field evolves. Corrections welcome.