Claude Platform GA: Computer Use, Browser Tool, Skills API, Files API

Anthropic's August 20, 2026 GA bundle is the production agent stack in one announcement: desktop control, in-browser control, your team's procedural knowledge, and document storage — no beta header on the new computer toolset.

Computer use, browser use, the Skills API, and the Files API are generally available on Claude Platform today, per Anthropic's launch post. The same week, Claude Academy opened for human upskilling — platform GA for agents, Academy GA for the humans steering them.

TL;DR — migration and architecture decisions

What shipped — four pieces one agent stack

Anthropic frames the bundle as operate software + apply expertise + return files:

Skills run in Anthropic's sandbox — nothing for you to host. Files let agents read uploaded documents once and write outputs you download later. That's the same separation explainx.ai teaches in loop engineering: harness owns execution, model owns reasoning.

Platform availability note from Anthropic: Skills API and Files API also ship on Microsoft Foundry; updated computer and browser tools are coming soon to Google Cloud Vertex AI. Existing beta integrations keep working during migration.

computer_toolset_20260801 — quick start

Add one entry to your Messages API tools array — no beta header:

{
  "model": "claude-opus-5",
  "max_tokens": 1024,
  "tools": [
    { "type": "computer_toolset_20260801" },
    { "type": "text_editor_20250728", "name": "str_replace_based_edit_tool" },
    { "type": "bash_20250124", "name": "bash" }
  ],
  "messages": [
    { "role": "user", "content": "Save a picture of a cat to my desktop." }
  ]
}

Supported models for the toolset include claude-fable-5, claude-mythos-5, claude-opus-5, claude-sonnet-5, and claude-opus-4-8, per compatibility docs.

Older models (Opus 4.7, Sonnet 4.6, etc.) remain on computer_20251124 with the beta header until you upgrade models and tool versions together.

The agent loop — how computer use actually runs

Anthropic's loop is explicit:

1. You send tools + a task requiring desktop interaction
2. Claude returns one or more tool_use blocks with stop_reason: "tool_use"
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 — that's the agent loop

If your integration only reads the first tool_use block, it will fail on the next API call once batch actions arrive. GA assumes full-turn handling.

Batch actions — ordered, halt-on-failure

A batch is several member calls in one assistant message — same JSON shape as parallel tool use, but sequential execution:

[
  { "name": "left_click", "toolset_name": "computer", "input": { "coordinate": [640, 60] } },
  { "name": "type", "toolset_name": "computer", "input": { "text": "pictures of cats" } },
  { "name": "screenshot", "toolset_name": "computer", "input": {} }
]

Rules that break naive loops:

• Run blocks in order — later actions depend on earlier ones
• On first failure, return is_error: true for that block
• For skipped later blocks, use 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 so it can verify state. You can attach a screenshot to the last result if it didn't — saves a round trip.

Migrate from computer_20251124 — nine-step checklist

Upgrading is optional until you're ready; computer_20251124 keeps working with its beta header. When you move:

Before/after shape change:

// Before (computer_20251124)
{ "name": "computer", "input": { "action": "left_click", "coordinate": [500, 300] } }

// After (toolset)
{ "name": "left_click", "toolset_name": "computer", "input": { "coordinate": [500, 300] } }

Anthropic's reference implementation remains the Docker-based starting point; GA docs point to a separate best-practices quickstart for production patterns (caching, compaction, trajectory recording).

Browser use — when not to spin up a desktop

For web-only tasks, browser use member tools target page elements rather than screen coordinates alone — closer to Playwright semantics without you hosting the browser stack in every integration.

Anthropic's guidance: declare both toolsets in one request if needed; they operate in independent coordinate frames, disambiguated by toolset_name. Compare to Claude in Chrome for the consumer-side browser story versus API-level browser use.

Skills API and Files API at GA

Skills API: Upload and version folder-based skills (instructions, scripts, templates) and attach them per request — same SKILL.md model as Claude Code filesystem skills, but API-managed and sandbox-executed. Pre-built skills (pptx, xlsx, docx, pdf) remain available by skill_id.

Files API: Upload once, reference by ID, download agent-written outputs. GA adds automatic expiration, 5× rate limits, and 1 TB per organization — meaningful for long computer-use trajectories that accumulate screenshots and documents.

Security — still your sandbox

GA doesn't remove risk. Anthropic repeats:

• Run in isolated VMs/containers with minimal privileges
• Avoid giving models sensitive credentials
• Allowlist internet access where possible
• Human confirmation for consequential actions
• Prompt injection via webpages/images remains a live threat — classifier defenses can be opted out via support for headless loops

Inform end users before enabling computer use in your product. Data retention: screenshots and inputs stay in your environment; API request retention follows Anthropic's standard policy — computer use is ZDR-eligible when you control storage.

What people are asking

Is this faster than before? Anthropic claims multi-action turns cut latency versus one-action-per-call loops — your mileage depends on batch size and screenshot payload. Image resizing and prompt caching matter more at scale.

Do I still need Codex for computer use? Our Codex vs Claude Code comparison noted Codex's native desktop control advantage — this GA closes the API-side gap for Claude Platform builders, not necessarily the Claude Code CLI experience overnight.

Managed Agents? Computer use is not currently available in Claude Managed Agents — build your own loop or wait for managed tooling.

The takeaway

August 20, 2026 is the day Anthropic treated desktop agents as production infrastructure: GA toolset, ordered batch actions, browser specialization, hosted skills, and file storage with org-scale limits. If you're on computer_20251124, the migration is mechanical but unforgiving — especially batch handling and screenshot sizing. If you're greenfield, start on computer_toolset_20260801 and pick browser use when the task never leaves the tab.

Related on explainx.ai:

• Codex vs Claude Code — Computer Use Comparison
• Claude Opus 5 Top 10 Use Cases — Computer Use
• What Are Agent Skills? Complete Guide
• How to Add Skills to Claude
• Loop Engineering for Coding Agents
• Claude Academy Launch
• Claude 5 Context Engineering
• Claude Cowork Record-a-Skill — Browser and Computer Use

Official sources: Anthropic GA announcement · Computer use tool docs · Agent Skills overview

Tool version strings, model compatibility, HIPAA eligibility, and Files API limits are accurate as of August 20–21, 2026 per Anthropic documentation. Vertex AI and regional rollouts may lag the Claude API GA date.