Claude Cowork on Web and Mobile

Claude Cowork, Anthropic’s agentic task environment, expanded from desktop-only to web and mobile as of July 7, 2026. The headline change is architectural: sessions now run remotely on Anthropic’s servers, so your sessions and files are tied to your Claude account rather than a single machine.

The expansion is in beta, rolling out first to Max plan holders over several weeks, with more plans to follow.

What shipped

Users on web (claude.ai) and mobile (iOS and Android, latest app versions) can now start, steer, and resume Cowork sessions from the same message box used for regular chat. Switching between Chat and Cowork modes happens inline via a toggle in the message box.

Because execution moved server-side, a session you begin on one surface can be picked up on another. The desktop app does not disappear. It becomes the bridge for capabilities that still require local access.

Where each feature works

The feature parity breakdown is the most operator-useful part of this announcement.

Capability	Availability
Start, steer, review tasks	All surfaces
Resume sessions across surfaces	All surfaces
Connectors, skills, plugins, projects	All surfaces
File preview, scheduled tasks	All surfaces
Local file access, browser use, computer use	Needs desktop as bridge
Live artifacts	Desktop only

When accessed from web or mobile, local file access, local connectors, browser use, and computer use work only while the Claude Desktop app is open on your machine. If the desktop app is closed, the remote session keeps running but loses access to local files and connectors.

Computer use remains a research preview, available only on Pro and Max plans. Projects tied to a local folder support Cowork sessions on desktop only. Adding content to a project is still a manual step, since Cowork will not modify a project’s contents automatically.

Why it matters

Cowork going cross-platform is less a product release and more an architectural commitment. The moment agent sessions become account-bound rather than device-bound, the UX contract changes. Agents stop being desktop-class features with mobile companion views and start behaving like ambient compute that follows the user.

Most agent platforms today still treat mobile as read-only or review-only. Cowork’s model, where a remote session surfaces identically on desktop, web, and mobile, is a direct answer to that gap. The desktop-as-bridge layer handles capabilities that genuinely need local access. That is a sensible layered architecture rather than a pretend parity model.

Because sessions now run server-side, scheduled tasks no longer require any device to be awake. Recurring and time-sensitive workflows run without leaving a machine on overnight.

How it compares

The “session runs on our servers, follows your account” model is not common among agent platforms today. Most either run locally, run in a sandboxed cloud environment with no cross-device continuity, or treat mobile as a read-only companion.

Grok is expanding agentic capabilities through Grok Build and a Voice Agent Builder (public beta as of July 1, 2026), with integrations into third-party platforms. That is a different bet: broad platform distribution rather than a unified account-bound session model. Neither approach is obviously correct. Anthropic’s model requires deeper investment in its own surface and infrastructure, while Grok’s model asks ecosystem partners to carry some of the reliability burden.

Cowork’s architecture is closer to what you would expect from a cloud IDE or a persistent compute service than from a chat product, which is the more accurate category comparison at this point.

Who should care

• Solo builders running scheduled automation get the most concrete win. Removing the “device must be awake” requirement means overnight workflows run reliably without a machine staying on.
• Mobile-primary operators can now join Cowork sessions from a phone instead of waiting until they reach a computer. Parity is not perfect, since live artifacts stay desktop-only, but the core task flow works.
• Platform engineers should map their dependency on local connectors and MCP servers before assuming web or mobile is a drop-in replacement. Any local-connector-dependent workflow still needs a desktop instance running somewhere.
• Founders placing a platform bet should note that Max gets this first, and the tier rollout beyond “more plans to follow” is unspecified. If your team is not on Max, the access timeline is genuinely unknown.

What to watch next

The rollout sequencing is worth monitoring. How quickly Anthropic extends Cowork to Pro and lower tiers will signal whether this is a differentiator they are protecting or infrastructure they want broadly adopted.

The desktop-as-bridge model is clever but fragile. A remote session that cannot reach local files when the desktop app is closed creates silent failures that are hard to debug: the task runs, the session continues, but a dependent file operation fails quietly. Expect operators to hit this, and expect Anthropic to need clearer failure signaling before the pattern is production-reliable.

Live artifacts remaining desktop-only is the most significant functional gap on web and mobile right now. If Anthropic ships live artifacts on web, that removes the last major reason a browser-based workflow would feel like a downgrade.

More broadly, the “sessions follow your account” model is the foundation for multi-device handoffs. Once stable, the second-order question is whether Cowork sessions become shareable or collaborative across accounts, not just across devices for one user.

Open questions

• The rollout timeline beyond Max is unspecified. “Several weeks” and “more plans to follow” leaves Pro and other tiers without a date to plan around.
• It is unclear what happens to an in-progress remote session if a user’s subscription changes mid-task. An edge case, but a real one for teams on billing cycles.
• There is no detail on remote session limits: maximum concurrent sessions, maximum duration, file storage caps, or behavior when limits are reached.
• The desktop-as-bridge failure mode is acknowledged but not fully specified. It is unclear how the disconnect is communicated, or whether the task errors out or pauses.
• “Research preview” for computer use is not defined in terms of reliability, feature completeness, or a path to general availability.

Source: support.claude.com