Most people still treat AI like a chat window. You ask, it answers, you close the tab.
Claude Cowork was built for a different contract: you hand Claude a goal, walk away, and come back to a draft deck, a cleaned spreadsheet, or a briefing doc waiting for review.
Until July 2026, that contract broke the moment you closed your laptop. Cowork lived on the desktop. The work stopped when you left the desk.
Anthropic's July 8 expansion to web and mobile fixes the operational half of that story. The bigger shift is architectural: sessions now run in Anthropic's cloud, bound to your account, not to one machine.
What actually changed on July 8
Three user-visible shifts from Anthropic's Cowork web and mobile post:
- Start on desktop, continue on phone. Check status, redirect, or approve from mobile.
- Background execution. Close the laptop mid-task; Claude keeps going.
- Scheduled tasks without a device online. Set Monday's client prep for 6am; it runs while every machine is asleep.
Anthropic also unified Chat and Cowork into one home on web and desktop so projects and artifacts live together.
Beta rolled to Max subscribers first, with other paid plans following. Usage limits were doubled through August 5, 2026 for launch.
Cloud sessions vs the old desktop-only model
The easy headline is "Cowork on your phone." The durable change is where execution runs.
Earlier Cowork required powered-on local hardware for many workflows. Remote sessions now run on Anthropic's servers, with files saved to your Claude account. That is why scheduled tasks can fire at 6am with no device online.
Anthropic's worked example is concrete: a morning task reads email threads and transcripts, pulls recent news, builds a briefing doc, and leaves a follow-up email drafted but unsent for human review over coffee.
That is agentic admin work, not a coding demo. It is also the workflow my ops clients ask for after they tire of copy-pasting CRM notes into chat.
Desktop is still the place for deep work when you need local folders and browser control. Web and mobile are the steering wheel for account-bound knowledge work.
Product page: Claude Cowork.
How this differs from recording Skills
I covered recording Skills in Cowork separately. Skills package repeatable UI workflows. Background cross-device execution is the runtime underneath:
| Feature | Skills recording | July 2026 cloud Cowork |
|---|---|---|
| Primary win | Reuse a narrated workflow | Keep working when you leave |
| Device binding | Tied to desktop capture | Account-bound cloud session |
| Scheduling | Yes, after packaging | Yes, including device-off runs |
| Best for | Repeatable ops chores | Long multi-step projects |
You will use both on the same account. Do not confuse "I can record a macro" with "my agent keeps running overnight."
When Claude pings your phone
Background agents fail in production when they hit a decision only a human can make. Anthropic's launch messaging includes mobile nudges when Claude needs approval instead of stopping cold.
That is the right default for client-facing agents:
- Auto-draft, human send on email
- Auto-research, human publish on external docs
- Auto-triage, human escalate on spend or legal flags
If you are designing voice or WhatsApp agents for clinics and service businesses, the same pattern applies: autonomy with an approval gate, not full autopilot on customer messages.
Surfaces and access in July 2026
| Surface | How to start Cowork | What you get |
|---|---|---|
| Web | Home screen at claude.ai | Cloud session, account files |
| Mobile | Sidebar in iOS/Android Claude app | Monitor, redirect, approve |
| Desktop | Full Cowork app | Local files + browser + cloud handoff |
Computer use (screen control) remains a last resort after connectors and browser tools. Anthropic asks permission per application when screen access is needed.
Help center: Claude Help Center for per-surface feature lists as beta expands.
What I would pilot first on a team rollout
If you are an operator, not a hobbyist, run one controlled pilot before you let everyone schedule overnight jobs:
- Monday briefing from email + calendar + one folder of notes. Human reviews output before send.
- Weekly CRM hygiene export with a fixed schema. No auto-write to production CRM until week three.
- Research pack for sales calls with explicit source list. Measure time saved vs hallucination cleanup.
Log where tasks start (desktop vs mobile) and how often humans get pinged. If phones buzz every ten minutes, your task definitions are too vague.
Limits I still plan around
Cloud Cowork does not replace:
- Local repo agents in Claude Code for git-heavy work
- Enterprise data residency reviews on cloud retention
- Fine-grained IAM for which connector can write vs read
It also does not solve evaluation. "Claude finished" is not "Claude was correct." Build review steps into every scheduled task the same way you would for API agents.
Bottom line
Cowork on web and mobile is not a cosmetic release. Cloud-bound sessions are what make background and scheduled agent work real for non-developers.
Close the laptop. Keep the task. Review on your phone. That is the product shape enterprise admins have been asking for since desktop agents went viral.
If you want help mapping Cowork-style background workflows onto your CRM, inbox, or client ops stack, book a free discovery call.