Claude's Python and TypeScript SDKs now ship built-in computer-use and browser-use toolsets: the SDK runs the agent loop, you only write the driver

Claude's computer-use tool only tells you what the model wants to click or type; until now every team wrote its own loop to map those actions to commands. Python SDK 1.12.0 and TypeScript SDK 0.132.0 (Oct 7, 2026) add toolsets: computer_toolset_20260801 and browser_toolset_20260801 are single tools[] entries that declare a family of member actions. You subclass an abstract toolset (for example BetaAbstractComputerToolset20260801) and implement only what your backend supports, whether VNC, xdotool, a DevTools client or a hosted browser; unimplemented members are sent as enabled: false. The tool_runner executes a turn's actions in order and stops at the first failure or refusal. The browser toolset runs your url_policy, file_policy and confirm hooks before each call, and a reference Chrome DevTools driver lives in claude-quickstarts. Anthropic published no new benchmarks for this SDK change. Before deploying, follow the seven-point safety checklist in browser-toolset.md: URL policy, request interception in the driver, container egress rules, a file policy or no file_upload, confirm gates for consequential actions, per-session isolation, and treating all page content as untrusted.