Claude's Background Computer Use Is Live, With a Commerce Agent Blueprint Alongside It

On September 2, Anthropic turned on background computer use for Claude Cowork and Claude Code: give Claude something to do on your desktop, and it clicks, types, and opens apps on its own while you work on something else entirely. On the same afternoon, Anthropic's developer team open-sourced Claude Commerce Agents, a public blueprint for building shopping and merchant agents. Neither is a small feature update. Together they mark the point where the useful skill with Claude stops being "write a good prompt" and starts being "decide what you're comfortable letting run while you're not watching."

What background computer use lets Claude do

Before this update, computer use meant Claude taking over your screen: you'd hand it a task, watch it move the mouse and type, and wait. Background computer use removes the waiting. On macOS 15 or later, Claude now runs in its own background window instead of commandeering your display, so you can keep writing, browsing, or coding in the foreground while it works a separate task behind the scenes. The same capability is available on Windows through the Claude Desktop app.

Under the hood, Claude follows a priority order when it needs to act: it reaches for a direct connector first (Slack, your calendar, whatever's already wired in), falls back to browser interaction next, and only resorts to raw screen clicking and typing as a last resort. That ordering matters more than it sounds. A connector call is a defined API request; a screen click is Claude guessing where a button is from a screenshot. The feature is available in Cowork and Claude Code, restricted to Pro and Max subscribers. Team and Enterprise plans don't have it yet.

The safety layer, and where it doesn't reach

Anthropic built a permission framework around the feature rather than shipping it wide open. Claude asks for approval before it touches a given application, and some categories are blocked by default: investment and trading platforms, cryptocurrency services. You can also set your own app blocklist to keep Claude out of anything specific, and it's worth doing that before you ever hand it a background task, not after.

The part worth reading twice is how Claude sees your screen at all: it navigates by taking screenshots, which means it can see anything visible on your desktop at that moment, including personal documents and anything else sitting in an open window. Anthropic's own guidance is to avoid granting access to banking, healthcare, or government sites for this reason. Claude is also trained to avoid stock trading, entering sensitive data, and collecting facial images, but Anthropic is explicit that these are trained behaviors, not hard technical enforcement. Treat them as a strong default, not a guarantee.

Turning it on is a Settings change: General, then the Cowork or Claude Code section, then the "Enable computer use" toggle. It's still in beta, and Anthropic's own notes flag that complex multi-step workflows sometimes need a retry, and that screen-based actions run slower than a connector would.

The other half: an open-source blueprint for shopping and merchant agents

The same day, Anthropic's developer team published Claude Commerce Agents on GitHub under Apache 2.0: a reference implementation for two kinds of agents, a shopping agent that searches, compares, and manages a cart for a customer, and a merchant agent that handles pricing, inventory alerts, and campaign drafts for the business side. It ships with working examples across four verticals, retail, travel, telecom, and entertainment, each wired to both a customer-facing storefront and a merchant portal.

The guardrails are the more interesting part for anyone weighing whether to build on it. Every dataset in the repo is fictional, no real orders are placed and no real cards are charged; checkout renders a cart for a human to complete manually, and every write a merchant agent makes is staged for approval rather than applied automatically. Anthropic's own docs/safety.md in the repo lists fencing and provenance gates on tool calls, operation caps, and grounding validation as enforced defaults across all three ways you can run it (the Messages API directly, the Claude Agent SDK, or a hosted Managed Agent). A Claude Code plugin can scaffold a new commerce agent against your own systems with a single /scaffold-commerce-agent command, and a companion /review-commerce-agent command audits one you already built. Anthropic is explicit that this is a reference implementation it doesn't maintain long-term and won't take contributions on, so treat it as a well-documented starting point, not a supported product.

What actually changed for anyone running agents

Neither announcement alone would be a big story. Together, they're a signal about where Anthropic thinks the interesting problems are now. Background computer use asks you to decide what Claude can touch while your attention is somewhere else. Commerce Agents asks the same question one layer further out: what should an agent be allowed to do on a customer's or a merchant's behalf without a human in the loop for every step. Both ship with the same instinct, real permission gates and staged writes rather than blanket trust, which is worth noticing given how much of the last year of agent releases skipped straight to "it can act now" without the accompanying "and here's exactly what it can't."

If you're already using Claude for anything unattended, the actual homework this week isn't trying the new feature. It's auditing your own blocklist and connector setup before you hand Claude a background task, the same way you'd check a new employee's access before their first day rather than after something goes wrong. For a broader look at where agent permissions and computer-use tooling stand across Anthropic's stack, see Claude's Computer Use Tool going generally available, the API-level building block this consumer feature builds on top of.

Sources: Anthropic Help Center, Let Claude use your computer in Cowork; @claudeai on X; @ClaudeDevs on Claude Commerce Agents; anthropics/commerce-agents on GitHub.