Claude Computer Use, Browser Use, Skills, and Files: GA Guide

Quick answer

Anthropic announced on August 20, 2026 that computer use, the new browser use tool, the Skills API, and the Files API are generally available on the Claude Platform. Together they can operate software, apply reusable procedures, read stored inputs, and return generated files.

GA does not turn them into one hosted agent product. Your application still runs the computer or browser tool calls in an environment it controls, decides which actions are authorized, returns every tool result, and owns recovery. Skills execute in Anthropic’s code-execution container. Files are workspace-scoped storage with separate expiration and deletion behavior.

What each surface owns

Surface	Use it for	Execution owner	Boundary to verify
Computer use	Full desktop applications and pixel-based UI work	Your application and desktop environment	OS isolation, credentials, screenshots, action approval
Browser use	Structured web interaction plus screenshot fallback	Your application and browser executor	Site allowlist, tabs, page content, prompt injection
Skills API	Versioned instructions, scripts, and templates	Anthropic code-execution container	Workspace sharing, version pinning, code review, runtime limits
Files API	Reusable input references and generated artifacts	Anthropic workspace storage	Access, expiration, deletion, audit trail, retention

Use browser rather than computer use when a task stays inside webpages. Use computer use when the agent must cross a browser, desktop application, and local UI. Use web search or web fetch instead when the task only reads public information and does not need an interactive page.

The agent loop remains your responsibility

Computer and browser use are client toolsets. Claude returns one or more ordered tool_use blocks; your application performs those actions and returns a matching result for each block. Later actions in a batch often depend on earlier actions, so run them in order and stop the remaining batch after the first failure.

Place authorization before execution, not after the whole batch. A single model turn can include a click, form value, and submission. Require fresh human confirmation before a payment, external message, account change, irreversible deletion, or scope expansion.

Set limits for iterations, elapsed time, cost, domains, destinations, files, and retries. A completion message is not evidence that the external state is correct. Capture a screenshot, receipt, file hash, API read-back, or another independent check.

Skills and files need lifecycle owners

Custom API Skills are shared across the workspace. Review every bundled instruction, script, template, and dependency as software. Pin a version in production, test it against a known-answer suite, and promote a new version only after reviewing its diff and generated artifacts. The Skills upload and versioning guide covers that lifecycle.

Files are also accessible to API keys in the same workspace. Treat file IDs as references, not authorization tokens. Set an expiration when the business purpose has a known end, delete superseded inputs, and test how expired references fail. The Files expiration and storage guide separates upload, reuse, download, expiration, and deletion.

A bounded production rollout

1. Pick one reversible workflow and record the exact model, toolset versions, skill versions, and workspace.
2. Use a dedicated browser or desktop sandbox with minimum data and privileges.
3. Define action-level approval, domain, file, cost, and iteration policies in code.
4. Upload only required files and set an explicit expiry where possible.
5. Run normal, denial, timeout, prompt-injection, duplicate-action, and recovery cases.
6. Verify external state independently and retain an audit record without leaking sensitive page content.
7. Start with a small cohort and rehearse revocation, rollback, file deletion, and skill-version rollback.

The production-readiness checklist turns these boundaries into a local review gate.

Provider and deployment boundaries

Anthropic’s launch post says the Skills API and Files API are also available through Microsoft Foundry, while updated computer and browser use are coming to Vertex AI. Do not infer identical timing, model compatibility, retention, pricing, regions, or controls across Claude Platform, Microsoft Foundry, Vertex AI, Amazon Bedrock, or Managed Agents.

Existing beta integrations can continue during migration, but production code should pin the current toolset and validate the response shape before switching. Current computer and browser documentation also says those client toolsets are not available in Claude Managed Agents.

Official sources

• Anthropic GA announcement
• Computer use documentation
• Browser use documentation
• Agent Skills overview
• Files API guide

Source check: August 23, 2026. Recheck toolset identifiers, supported models, deployment availability, retention eligibility, and live account access before implementation.