Claude requests a tool call, and the documentation says who runs it

The conversation

Anthropic documents tool use, also called function calling, as a way for Claude to call functions an application defines or that Anthropic itself provides. As Anthropic's current API reference, retrieved 16 September 2026, puts it, 'Claude determines when to call a tool based on the user's request and the tool's description,' then 'returns a structured call that your application executes (client tools) or that Anthropic executes (server tools).' Anthropic's own announcement of general availability, dated 30 May 2024, made tool use available across the Claude 3 model family on Anthropic's API, Amazon Bedrock and Google Cloud's Vertex AI, citing uses such as extracting structured data from documents and answering questions by querying a database.

What the documents show

The reference describes a defined round trip for a client tool: a request lists a tool's name, description and JSON input schema; Claude's response carries a stop reason of 'tool_use' with a block naming the tool and its arguments; the calling application runs the operation and sends the outcome back inside a 'tool_result' block; and Claude uses that result to produce its answer. The May 2024 announcement lists example uses including document processing and API orchestration, framed throughout as what a developer's own code does with Claude's structured output rather than as Claude acting directly on outside systems.

The system boundary

The documentation draws an explicit line between two categories. For a 'client tool,' including one a developer writes or an Anthropic-schema tool such as its bash or text editor tool, the reference states plainly that 'your application executes the operation.' For a 'server tool,' such as web search or code execution, Anthropic's own infrastructure performs the call and Claude receives the result directly. Either way, the documentation is explicit that Claude itself does not reach outside the boundary of a single request-response turn without a system, the calling application or Anthropic's own servers carrying out the actual operation.

Where it fails

Because Claude decides whether a request matches a tool's description well enough to call it, the documentation itself flags a specific limitation: when a required parameter is missing from a user's prompt, the reference notes a model 'might infer a reasonable value' rather than asking for it, behavior it calls 'not guaranteed' and more likely with smaller models on ambiguous prompts. A builder relying on tool use to gate a real-world action inherits that uncertainty about when the model asks rather than assumes.

• Does the application validate a tool's arguments before executing an action with real consequences?
• Is a missing or inferred parameter ever passed through to a live system without confirmation?
• For a server tool executed on Anthropic's infrastructure, what does the application still need to log to reconstruct what happened?

Anthropic's own documentation is consistent that a model requests an action; something else, inside a stated boundary, performs it.