update · TowCue Editorial Team
Anthropic Computer Use, Browser Use, Skills API & Files API reach GA
Anthropic moved Computer Use, Browser Use, Skills API, and Files API to general availability on August 20, 2026. Batch actions, element-level browser control, hosted skills, and managed file storage — all beta headers dropped. What changes for production agent builders.
Quick answer
Anthropic announced general availability (GA) for four agent infrastructure components on August 20, 2026:
| Component | What's new at GA | Toolset ID |
|---|---|---|
| Computer Use | Batch multi-action turns, no beta header, HIPAA-eligible (BAA), zoom enabled by default | computer_toolset_20260801 |
| Browser Use (new) | Element-level web interaction via accessibility tree, runs in developer's browser environment | browser_toolset_20260801 |
| Skills API | Upload/version custom skills, attach per request, sandbox execution, pre-built skills (pptx, xlsx, docx, pdf) | skills-2025-10-02 header dropped |
| Files API | 1 TB/org storage, 5× rate limits, auto-expiration (1 hr–90 days), reference by ID | files-api-2025-04-14 header dropped |
All beta headers dropped. Old beta versions (computer_20251124, skills-2025-10-02, files-api-2025-04-14) continue working during migration.
Availability (as of Aug 20): Both computer-facing tools work on five models only (claude-opus-5, claude-sonnet-5, claude-fable-5, claude-mythos-5, claude-opus-4-8) and only on the Claude API — not Bedrock, Vertex AI, Microsoft Foundry, or Claude Managed Agents. Updated versions coming soon to Vertex AI.
Official source: Anthropic Blog — "Build production agents with computer use, the Skills API, and the Files API" (August 20, 2026)
Platform docs: Computer Use Tool | Computer Use Tool (MD)
What changed at GA
1. Computer Use: batch actions, fewer round trips
The headline mechanical change: batch actions. The new computer_toolset_20260801 lets Claude return multiple member calls in one turn (e.g., click → type → screenshot), executed strictly in order with halt-on-failure.
Before (beta computer_20251124) | After (GA computer_toolset_20260801) |
|---|---|
| One action per model turn | Multiple actions per turn (batch) |
| Fresh model call between each action | Single turn, sequential execution |
| Beta header required | No beta header |
| No zoom by default | Zoom enabled by default |
Parameter: display_width_px, display_height_px, enable_zoom | Single configs object for per-tool settings |
| HIPAA: not eligible | HIPAA-eligible (BAA) |
Migration is not a version bump. Parameters display_width_px, display_height_px, enable_zoom are removed — replaced by a configs object. Existing integrations need code changes (see 9-step migration checklist).
Supported models (GA): claude-opus-5, claude-sonnet-5, claude-fable-5, claude-mythos-5, claude-opus-4-8. Older models stay on computer_20251124 with beta header.
Platforms: Claude API only. Bedrock, Vertex AI, Microsoft Foundry, Claude Managed Agents — not yet supported for the new toolset (Vertex AI "coming soon").
2. Browser Use: new toolset, element-level control
A genuinely new capability: browser_toolset_20260801 (27 default member tools, 31 total). Unlike Computer Use which controls a full desktop via screenshots/coordinates, Browser Use:
- Runs inside your application's browser viewport (not Anthropic's infra)
- Reads the page's accessibility tree alongside screenshots
- Targets elements by stable references (
ref_4for "Submit button") instead of screen coordinates - References survive layout shifts and work across tabs
- 27 default tools: navigation, clicks, typing, form input, tab management, download reporting
- 4 optional tools (explicit enable required): file upload, JavaScript execution, 2 diagnostic tools
When to prefer Browser Use: Tasks that never leave the browser — form filling, web app workflows, SaaS automation. Closer to Playwright semantics without hosting the browser stack yourself.
When to prefer Computer Use: Cross-application workflows, desktop apps, tasks needing OS-level interaction.
Declare both in one request if needed — they operate in independent coordinate frames, disambiguated by toolset_name.
3. Skills API: hosted procedural knowledge
What changed: Beta header skills-2025-10-02 dropped. Mechanics otherwise unchanged.
- Skill = folder with
SKILL.md(YAML frontmatter: name ≤64 chars, description ≤1024 chars), instructions, scripts, templates - Under 30 MB uncompressed, code execution tool must be enabled to run
- Up to 20 skills per request, mix Anthropic pre-built (
pptx,xlsx,docx,pdf) with custom uploaded/versioned skills - Loaded via
containerparameter on Messages API requests - Runs in Anthropic's code execution sandbox — nothing to host yourself
Use case: Encode team procedures (e.g., "how we process invoices," "our PR review checklist") once, attach to any agent request that needs them.
4. Files API: managed storage with real limits
What changed: Beta header files-api-2025-04-14 dropped. New GA features:
| Feature | GA spec |
|---|---|
| Storage | 1 TB per organization |
| Rate limits | ~500 file-related requests/minute (5× beta) |
| File size | 500 MB max per file |
| Auto-expiration | Configurable at upload: 1 hour – 90 days (expires_in_seconds) |
| Reference | Upload once, reference by ID in later requests, download agent outputs |
Use case: Long computer-use trajectories accumulate screenshots/documents — upload once, reference by ID, let auto-expiration clean up.
What this means for agent builders
The agent loop is now explicit and unforgiving
GA assumes full-turn handling of the agent loop:
1. You send tools + task requiring desktop/browser interaction
2. Claude returns one or more `tool_use` blocks (batch)
3. Your app executes each action in your sandboxed environment
4. You return matching `tool_result` blocks (images for screenshot/zoom, short text for clicks/types)
5. Repeat until Claude responds with text (task completion)
Critical: If your integration only reads the first tool_use block, it will fail on the next API call once batch actions arrive. GA assumes you loop over every block.
Batch halt semantics are mandatory
| Rule | Implementation |
|---|---|
| Run blocks in order | Later actions depend on earlier ones |
| On first failure | Return is_error: true for that block |
| For skipped later blocks | Return exactly: "Not executed: an earlier computer action in this turn failed." |
Every tool_result | Must echo "toolset_name": "computer" |
| Leaving any block unanswered | Triggers invalid_request_error |
Claude often ends batches with screenshot to verify state. You can attach a screenshot to the last result if it didn't — saves a round trip.
Screenshot sizing: strict at GA
The new toolset rejects oversized images instead of downscaling. You must resize oversized screenshots before returning them (step 9 of migration checklist).
Consumer side: Claude in Chrome GA (Aug 26)
Separate but coordinated: Claude in Chrome reached GA on August 26, 2026 (Anthropic Blog).
- Available on every paid Claude plan
- Auto-approval for actions a safety classifier deems safe (toggleable in settings)
- Classifier validates each action against original request before execution
- Prompt injection defenses hardened after pilot period
- Enterprise admins can manage via Organization settings
Claude Cowork built-in browser (separate Chromium instance, sandboxed side panel) also shipped Aug 26 for Enterprise, rolling out to Pro/Max/Team that week — completely isolated from user's browser session, credentials imported site-by-site.
Limitations and gaps
| Gap | Status |
|---|---|
| Managed Agents | Computer Use not available in Claude Managed Agents — build your own loop or wait |
| Bedrock / Vertex AI / Foundry | Updated computer/browser tools not yet available (Vertex AI "coming soon") |
| Older models | Stay on computer_20251124 with beta header until model + tool upgrade together |
| Managed Agents API domain allowlists | Added Aug 19 for Managed Agents; not yet extended to built-in browser |
| Enterprise credential import allowlist | Not described for built-in browser in current announcement |
TowCue take
August 20, 2026 is the day desktop agents became production infrastructure. The GA bundle — ordered batch actions, browser specialization, hosted skills, org-scale file storage — is the first time a major provider has shipped a complete, no-beta-header agent stack for desktop automation.
For teams building automation products: The migration from computer_20251124 is mechanical but unforgiving (batch handling, screenshot sizing, configs object). If you're greenfield, start on computer_toolset_20260801 and pick Browser Use when the task never leaves the tab.
For enterprise buyers: Computer Use is now HIPAA-eligible (BAA). The 1 TB Files API, 5× rate limits, and auto-expiration make long trajectories viable. But the platform gap (Claude API only, no Managed Agents, no Bedrock/Vertex/Foundry for new toolset) means lock-in to Anthropic's API for now.
For governance: The EU AI Act (generally applicable Aug 2, 2026) and SR 26-2 (April 2026) require logging, traceability, human oversight for high-risk agentic AI. A signed, independently verifiable receipt at the admission boundary (before execution) is the concrete answer — not post-hoc logs. Build that into your loop now.
Verify current model/platform support on the Computer Use Tool docs before committing to an integration path.
Research sources
- Anthropic Blog — "Build production agents with computer use, the Skills API, and the Files API" (August 20, 2026)
- Anthropic Platform Docs — Computer Use Tool (current)
- Anthropic Platform Docs — Computer Use Tool (MD) (migration checklist, agent loop, batch semantics)
- Anthropic Blog — "Claude in Chrome is generally available" (August 26, 2026)
- Anthropic Help Center — Let Claude use your computer in Cowork (current limitations, Pro/Max only)