Claude 工具使用
Claude 能够与工具和函数交互,使您能够扩展 Claude 的能力以执行更多样化的任务。
以下是如何使用 Messages API 向 Claude 提供工具的示例:
工具使用的工作原理
Claude 支持两种类型的工具:
-
客户端工具:在您的系统上执行的工具,包括:
- 用户定义的自定义工具,由您创建和实现
- Anthropic 定义的工具,如计算机使用(测试版)和文本编辑器,需要客户端实现
-
服务器工具:在 Anthropic 服务器上执行的工具,如网络搜索工具。这些工具必须在 API 请求中指定,但不需要您自己实现。
Anthropic 定义的工具使用版本化类型(例如,web_search_20250305
,text_editor_20250124
)以确保跨模型版本的兼容性。
客户端工具
通过以下步骤将客户端工具与 Claude 集成:
向 Claude 提供工具和用户提示
- 在 API 请求中定义客户端工具,包括名称、描述和输入模式。
- 包含可能需要这些工具的用户提示,例如,“旧金山的天气如何?”
Claude 决定使用工具
- Claude 评估是否有任何工具可以帮助解答用户的查询。
- 如果是,Claude 构建一个格式正确的工具使用请求。
- 对于客户端工具,API 响应的
stop_reason
为tool_use
,表示 Claude 的意图。
执行工具并返回结果
- 从 Claude 的请求中提取工具名称和输入
- 在您的系统上执行工具代码
- 在包含
tool_result
内容块的新user
消息中返回结果
Claude 使用工具结果制定响应
- Claude 分析工具结果,为原始用户提示制作最终响应。
注意:步骤 3 和 4 是可选的。对于某些工作流程,Claude 的工具使用请求(步骤 2)可能就是您所需的全部内容,无需将结果发送回 Claude。
服务器工具
服务器工具遵循不同的工作流程:
向 Claude 提供工具和用户提示
- 服务器工具,如网络搜索,有自己的参数。
- 包含可能需要这些工具的用户提示,例如,“搜索有关 AI 的最新新闻。”
Claude 执行服务器工具
- Claude 评估服务器工具是否可以帮助解答用户的查询。
- 如果是,Claude 执行该工具,结果会自动纳入 Claude 的响应中。
Claude 使用服务器工具结果制定响应
- Claude 分析服务器工具结果,为原始用户提示制作最终响应。
- 服务器工具执行不需要额外的用户交互。
工具使用示例
以下是演示各种工具使用模式和技术的几个代码示例。为简洁起见,这些工具都是简单工具,工具描述比理想情况下确保最佳性能的描述要短。
定价
Tool use requests are priced based on:
- The total number of input tokens sent to the model (including in the
tools
parameter) - The number of output tokens generated
- For server-side tools, additional usage-based pricing (e.g., web search charges per search performed)
Client-side tools are priced the same as any other Claude API request, while server-side tools may incur additional charges based on their specific usage.
The additional tokens from tool use come from:
- The
tools
parameter in API requests (tool names, descriptions, and schemas) tool_use
content blocks in API requests and responsestool_result
content blocks in API requests
When you use tools
, we also automatically include a special system prompt for the model which enables tool use. The number of tool use tokens required for each model are listed below (excluding the additional tokens listed above). Note that the table assumes at least 1 tool is provided. If no tools
are provided, then a tool choice of none
uses 0 additional system prompt tokens.
Model | Tool choice | Tool use system prompt token count |
---|---|---|
Claude Opus 4 | auto , none any , tool | 346 tokens 313 tokens |
Claude Sonnet 4 | auto , none any , tool | 346 tokens 313 tokens |
Claude Sonnet 3.7 | auto , none any , tool | 346 tokens 313 tokens |
Claude Sonnet 3.5 (Oct) | auto , none any , tool | 346 tokens 313 tokens |
Claude Sonnet 3.5 (June) | auto , none any , tool | 294 tokens 261 tokens |
Claude Haiku 3.5 | auto , none any , tool | 264 tokens 340 tokens |
Claude Opus 3 | auto , none any , tool | 530 tokens 281 tokens |
Claude Sonnet 3 | auto , none any , tool | 159 tokens 235 tokens |
Claude Haiku 3 | auto , none any , tool | 264 tokens 340 tokens |
These token counts are added to your normal input and output tokens to calculate the total cost of a request.
请参阅我们的模型概述表了解当前每个模型的价格。
当您发送工具使用提示时,就像任何其他 API 请求一样,响应将输出输入和输出令牌计数,作为报告的 usage
指标的一部分。
下一步
在我们的 cookbook 中探索随时可以实现的工具使用代码示例: