Tool Use (Function Calling)
Summary — Key Takeaways from This Lesson
- Tool Use is the mechanism that allows Claude to call user-defined external functions. The fundamental division of responsibility is: Claude decides, your app executes.
- Tools are defined using JSON Schema. Three fields are required:
name,description, andinput_schema. - When Claude returns a
tool_useresponse, your app executes the function and sends the result back as atool_result. This back-and-forth is called the agentic loop. - Server-side tools provided by Anthropic (such as
web_search) are executed on Anthropic's side, so no agentic loop implementation is required. - The code examples here are based on the official documentation as the primary source.
目次 (6)
How Tool Use Works
Claude is a language model specialized in text generation. However, with Tool Use, Claude can "request" real-time data retrieval, computation, and operations on external services. The actual processing is handled by your app, while Claude's role is to decide what to call and when source.
Defining Tools (JSON Schema)
Tools are passed as an array in the tools parameter.
The input_schema for each tool is written in JSON Schema format
source.
# Example tool definition (Python)
tools = [
{
"name": "get_weather",
"description": "指定した都市の現在の天気を取得します。",
"input_schema": {
"type": "object",
"properties": {
"location": {
"type": "string",
"description": "都市名。例: Tokyo, Japan"
},
"unit": {
"type": "string",
"enum": ["celsius", "fahrenheit"],
"description": "温度単位"
}
},
"required": ["location"]
}
}
]
Python — Minimal Code Example
The following is based on the code example from the official documentation source.
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(
model="claude-opus-4-7",
max_tokens=1024,
tools=tools, # 上記で定義したツール
messages=[{"role": "user", "content": "東京の天気を教えてください"}]
)
# stop_reason が "tool_use" のとき Claude がツール呼び出しを要求している
print(response.stop_reason) # => "tool_use"
print(response.content) # => tool_use ブロックを含む配列
The Agentic Loop Flow
When using Tool Use, the basic sequence of interactions is as follows.
- Your app sends a request to Claude along with the tool definitions
- Claude returns
stop_reason: "tool_use"along with atool_useblock - Your app reads the
nameandinputfrom thetool_useblock and executes the corresponding function - The execution result is added to the message array as a
tool_resultand sent to the API again - Claude generates and returns the final answer text
For detailed implementation steps, refer to the official documentation.
Controlling Tool Use with tool_choice
The tool_choice parameter lets you control how Claude selects tools.
"auto"(default) — Claude decides whether to use a tool"any"— Claude must use one of the provided tools"tool"+"name"— Claude must use the specified tool"none"— Claude does not use any tools
Tips for Designing Tools
The description field is the primary basis Claude uses to decide when to use a tool.
Writing 3–4 or more sentences clearly explaining "what it does," "when to use it," and "what it returns"
significantly improves accuracy
source.
