Anthropic Claude Tool Use

What tool use actually enables

Claude's tool use API allows Claude to call external functions, APIs, and services as intermediate steps toward completing a task — rather than generating a response from training data alone. A Claude agent equipped with tool use can query a database, update a CRM record, trigger an external workflow, or retrieve a live document mid-task. Anthropic's tool use documentation notes that on benchmarks like SWE-bench (real-world software engineering) and LAB-Bench FigQA (scientific figure interpretation), adding even basic tools produces outsized capability gains — often surpassing human expert baselines.

The core loop: Claude receives a task, decides a tool needs to be called, returns a structured tool_use block with the parameters, your application (or Anthropic's infrastructure) executes the function, and the result is returned to Claude as context for the next step. Multiple tool calls can chain within a single response, enabling multi-step agent workflows.

Client tools vs server tools

Tools divide into two execution categories. Client tools — including user-defined functions and Anthropic-schema tools like bash and text_editor — run in your application. Claude responds with stop_reason: 'tool_use', your code executes the operation, and you return the result. Server tools — web_search, code_execution, web_fetch — run on Anthropic's infrastructure. You see results directly without managing execution, which removes the infrastructure overhead for common operations like web search.

The choice between client and server tools matters for production architectures. Client tools give you full control over execution environment, authentication, and data handling. Server tools are faster to implement but require trusting Anthropic's infrastructure with the query and result.

Strict mode and production reliability

Two features have significantly improved production reliability for tool use. First, strict mode: adding strict: true to tool definitions forces Claude's tool calls to always match the defined schema exactly — eliminating the class of errors where Claude produces a tool_use block with missing or incorrectly typed parameters. Second, fine-grained tool streaming, available on Claude Sonnet 4.5, Haiku 4.5, and Opus 4, reduces latency by streaming tool use parameters without buffering — relevant for workflows where tool results feed directly into subsequent reasoning steps. The full tool reference documents both features with implementation examples.

Nuevexa builds AI agents using Claude's tool use API, integrating with CRM, billing, project management, and custom business systems. Our AI Agents service covers the full architecture from tool definition through production deployment and monitoring.