Tool use

Tool use is the base mechanism the whole agent stack is built on. You describe tools; Claude decides when to call them; you return results; Claude continues.

Client-side vs server-side

Client-side tools are yours. You define the schema, Claude asks for a call, your code executes it. Priced like any other request — you pay for the tokens.

Server-side tools are hosted by Anthropic and execute without a round trip to you. These can carry additional usage-based charges:

Server tool	Additional cost
Web search	$10 per 1,000 searches, plus tokens for the content retrieved
Web fetch	None beyond standard token costs
Code execution	1,550 free hours per organisation per month, then $0.05 per container-hour. Free when used alongside web search or web fetch.
Bash, text editor	Token cost only
Computer use, browser use	Token cost only, but the toolset definitions are large — see below

The cost nobody budgets for: tool definitions

Declaring tools adds tokens to every request, before anything is called:

What	Approximate input tokens
Tool-use system prompt (Opus 5, auto/none)	286
Tool-use system prompt (Opus 5, any/tool)	406
Bash tool definition (Opus 5)	325
Text editor tool (text_editor_20250429)	700
Computer use toolset (computer_toolset_20260801)	~4,500
Browser use toolset (browser_toolset_20260801)	~6,600

A browser-use agent therefore starts every single request roughly 7,000 input tokens in the hole. Prompt caching is not an optimisation there — it is the difference between a viable and an unviable cost model.

Note the counts differ per model: the same auto tool choice costs 286 tokens on Opus 5 and 675 on Opus 4.7. Do not carry an old estimate across a model upgrade.

What it does not do

• It does not execute your tools. Client-side means client-side; Claude only asks.
• It is not MCP. MCP is a protocol for distributing tools. Tool use is the underlying model capability MCP servers plug into.
• Errors are not always billed. A failed web search is not billed; a failed tool call still consumed tokens.