Computer use — letting an AI model click, type, and navigate a real desktop instead of calling an API — has always had one practical limitation: it takes over your screen. You start a task, hand over the mouse, and wait. In early September 2026, Anthropic shipped background computer use for Claude Cowork and Claude Code, removing that constraint for macOS users on Pro and Max plans.
This isn't a new model or a new capability in the abstract sense. It's a scheduling and UX change, and those are often the updates that actually shift how teams work day to day.
What Changed
Previously, computer use meant your machine was unavailable while Claude operated it — the same experience as remote-controlling your own laptop. With background execution, Claude opens and drives apps in background windows while you keep typing, browsing, or working in the foreground. It won't grab your mouse or keyboard mid-task, and it pauses if you're actively typing. The only time it asks for the full screen is when a task genuinely requires it, and even then it asks permission once per session rather than repeatedly.
The setting lives under Settings → General → Computer use, and it's opt-in — off by default.
Where This Fits an Agentic Workflow
At webhani, we think about agentic tooling in terms of what it removes from a human's critical path, not just what it automates. Background computer use targets a specific gap: tasks that have no API to call. Internal admin dashboards, legacy enterprise tools, mobile app simulators for manual QA passes — these are common in client environments and are exactly the kind of work that previously required a human to sit and click through screens.
A realistic pattern looks like this: a developer kicks off a background task ("walk through this internal CRM and export the last quarter's account list to CSV") and continues working on a feature branch in the foreground. The two tasks don't compete for screen focus.
A Conceptual Permission Pattern
Claude Code's settings.json already uses an allow/deny permission model for tool access. Enabling any autonomous or background-capable tool should go through the same explicit, scoped approval rather than a blanket grant:
{
"permissions": {
"allow": [
"ComputerUse(app:internal-crm)"
],
"deny": [
"ComputerUse(app:*)"
]
}
}Treat this as illustrative rather than a literal schema — always check Anthropic's current documentation for the exact permission keys before configuring it, since these interfaces evolve quickly.
Risk Considerations
Background automation raises the stakes on oversight, not just convenience. Three things we'd insist on before enabling it on a client machine:
- Scope tightly. Grant access to specific applications, not the whole desktop.
- Use a dedicated profile or machine for background automation, separate from anything holding production credentials or unrelated sensitive sessions.
- Review the action log after every run. Background tasks are easy to fire and forget — don't let them run unattended for long stretches without a human checking what actually happened.
Practical Use Cases for Client Teams
- Compiling competitive research from local files while a developer keeps coding.
- Running a manual QA pass in a mobile simulator to catch UX regressions an automated test suite wouldn't flag.
- Navigating internal dashboards or specialized enterprise software that has no formal API — common in legal, healthcare, and finance clients we work with.
webhani's Recommendation
We'd treat this as a targeted tool for API-less, repetitive desktop work — not a replacement for proper automation where an API already exists. If a task can be scripted or hit through an SDK, that's still the more reliable, auditable path. Background computer use earns its place specifically where no such path exists.
Start with a single low-risk workflow on a sandboxed machine, watch the logs for a week, and only then consider expanding scope.
Takeaways
Background computer use is a UX shift more than a capability leap, but UX shifts are often what determine whether a feature actually gets used. For teams doing agentic development, it closes a real gap around API-less enterprise tools — as long as it's rolled out with the same scoping discipline you'd apply to any tool with desktop-level access.