The Python and TypeScript SDKs added beta classes for Claude Browser Use and Claude Computer Use: you subclass one and write one method per tool against your own browser or desktop automation.
This is a small release note with a larger meaning. Until now every integration wrote the same dispatch loop — receive tool calls, map each to Playwright or a VNC session, marshal results back in the right shape. That code was boilerplate, and boilerplate everybody writes separately is boilerplate everybody gets subtly wrong, particularly around batched actions and error reporting.
It does not change what the agent can do, and it changes nothing in the table. It is recorded because the gap between a model can emit click actions and a team can ship this is mostly made of this kind of work, and that gap is narrowing faster than the capability itself.