Anthropic makes Computer Use, Browser Use, Skills API, and Files API generally available

Anthropic moved four agent-building primitives to general availability on August 20, 2026: agents can operate software, load versioned skills, and exchange persistent files through the Claude Platform.

Anthropic made Computer Use, the new Browser Use tool, Skills API, and Files API generally available on the Claude Platform on August 20, 2026. The package targets production agents that can operate software, load an organization’s procedures, and return finished files instead of stopping at a chat response. (Source: Anthropic, 2026-08-20)

Key facts:

Computer Use can batch several actions per model turn, reducing the number of model calls in a workflow.
Browser Use adds page structure to screenshots, so an agent can target a web field or button instead of relying only on pixel coordinates.
Skills API uploads and versions folders of instructions, scripts, and templates. Skills run in Claude’s code-execution sandbox.
Files API supports files up to 500 MB and 1 TB of storage per organization. Files can also expire automatically between 1 hour and 90 days.
Skills API and Files API are available through Microsoft Foundry. The updated computer-use tools are coming soon to Google Cloud Vertex AI.

Anthropic illustration for Computer Use, Skills API, and Files API — Anthropic’s official launch artwork for the four production-agent building blocks. (Source: Anthropic)

What changed for builders

The practical change is a cleaner agent loop. Computer Use handles applications that do not expose an API: Claude receives a screenshot, then clicks, types, and scrolls. Browser Use is the web-specific addition; it combines the visual page with structural information, which should make web automation less brittle than hard-coded coordinates. (Source: Computer Use documentation)

Skills are the reusable procedure layer. A team can package domain rules, scripts, templates, and reference material as a versioned folder, then attach that skill to requests that need it. Anthropic describes progressive disclosure: metadata is available for discovery first, while detailed instructions are loaded only when the task needs them. That makes a skill closer to a maintained operating procedure than to a long system prompt. (Source: Agent Skills documentation)

Files are the state layer. An uploaded PDF or spreadsheet can be referenced by file ID in later requests instead of being sent repeatedly. Generated files from Skills or code execution can be downloaded, while uploaded files are not downloadable through the API. The Files API also supports expiration, which is useful for temporary customer documents and reduces manual cleanup. (Source: Files API documentation)

Anthropic uses Box Agent as an example of applying specialized document-generation skills to files already stored in Box. (Source: Anthropic)

The builder takeaway

This is most useful for workflows with three missing pieces: a legacy web application with no API, internal know-how that must be applied consistently, and documents that need to persist across steps. A claims workflow, for example, can read an intake file, apply a filing skill, complete a web portal submission, and save the confirmation as a new file.

The security boundary still matters. Anthropic’s documentation recommends isolated environments, limited network access, and human confirmation before consequential actions. Browser Use makes agents more capable; it does not remove prompt-injection risk from the pages they read.

Existing beta integrations continue to work while teams migrate. For a new build, the sensible first test is a narrow workflow with expiring files, a read-only or sandboxed browser target, and a skill whose output can be checked against a fixed rubric.

Sources: Anthropic announcement · Computer Use documentation · Agent Skills documentation · Files API documentation