Article

Claude Cowork Moves New Pro and Max Tasks to the Cloud

What Claude Cowork’s October 6 Pro and Max cloud change means for local files, scheduled tasks, Claude Code and organizational controls.

Editorial illustration for Claude Cowork Moves New Pro and Max Tasks to the Cloud: a geometric block represents a model release. Not documentary evidence.

Anthropic’s October 6, 2026 change moves new Claude Cowork tasks on Pro and Max plans to the cloud and removes the “Only on your computer” setting. Existing local tasks remain on their original computer. For subscribers using files or recurring workflows, the immediate decision is which tasks can continue without a connected desktop. Anthropic’s getting-started guide documents the change.

October 6 is the stated effective date, not a verified first-announcement timestamp. The broader merger of Cowork and chat was announced on September 16 and is rolling out gradually; some accounts no longer show separate Chat and Cowork choices.

Cloud execution still has desktop dependencies

The agent runs in Anthropic’s cloud, while access to local files still depends on the desktop app. Browser operation and screen control also need Claude Desktop open and online on macOS or Windows. Cloud and web/mobile support remain labeled beta. These are documented requirements, not a guarantee that every workflow can finish unattended.

For connected-folder access, the migration guide requires a session started on desktop and that app to remain open. Anthropic says it copies individual files into the cloud as needed, rather than uploading entire connected folders. Local MCP connectors also depend on Desktop. Closing the app cuts off those device resources, even if the hosted session continues.

This decision framework separates where the agent runs from what each step needs. The choices are analysis of the linked documentation.

Your requirement

Practical choice

Condition to check

Continue knowledge work across devices

Cloud Cowork

Use approved account-stored inputs or cloud connectors; sessions follow the account.

Work with files, browser actions or screen control on one computer

Cowork with a connected desktop

Keep the machine and Desktop available for those steps. Desktop requirements.

Run suitable development tools on your own machine

Claude Code with Local selected

Local tool execution still sends context to a model provider. Code environments.

Keep all model-visible information offline

Neither documented consumer path establishes this

Local Code still exchanges prompts and outputs over the network. Data-flow documentation.

Apply organization-wide access and monitoring rules

Evaluate Team or Enterprise controls

Check actual cloud enablement, role grants and logging configuration. Managed-plan controls.

Treat local-file schedules as unresolved

The migration notice says existing schedules move to the cloud, including local-file tasks when Desktop is open. But the scheduling guide says schedules cannot be attached to a local folder, while its manual-setup section says tasks requiring local files or apps run locally. The documentation does not clearly reconcile these statements.

Do not assume a local-folder schedule will migrate unchanged. Review its instructions, dependencies and next run in the Scheduled view, and seek product clarification before relying on it. The guide supports schedules using account files and cloud connectors without a device online. That does not settle local-folder support or migration fidelity.

Illustrative dependency check: a briefing built entirely from approved cloud inputs has different requirements from one that must update a workbook on a laptop. The latter still needs access to that laptop at the write step. Moving the agent’s runtime does not move every dependency; this is a planning distinction, not a test result.

Claude Code offers local execution, not offline inference

For suitable machine-bound work, Claude Code Desktop offers Local, Cloud and SSH environments. Select Local deliberately rather than assuming the desktop interface determines execution location. Local and SSH sessions send conversations and code context to the configured model provider, Anthropic’s API by default.

Anthropic provides transcript and history downloads for continuing Cowork work in Code, but projects and schedules do not transfer. Treat that as context you can bring into another tool, not a complete workspace migration.

This can address a requirement to run tools on your machine. It does not establish that model-visible data stays there. The same distinction matters for organizations considering self-hosted Claude Code execution; execution location and inference location need separate answers.

Check file access and retention separately

Anthropic describes temporary, isolated cloud sandboxes for each session, with outbound sandbox traffic passing through an allowlist proxy. That describes execution isolation, not conversation deletion or permission to process sensitive information. These are vendor architecture claims, not an independent security assessment.

Under the consumer retention policy, deleting a conversation removes it from history immediately and from backend storage within 30 days. If model improvement is enabled, de-identified data may remain in training pipelines for up to five years; policy and legal exceptions also apply. Check the model-improvement setting separately from file permissions. Ending a sandbox is not the same as deleting stored conversation data.

The Cowork safety guide also distinguishes sandbox isolation from actions through connected apps. Untrusted content can still introduce prompt injection, and granted tools can affect real accounts or files. Restrict working folders, retain backups and supervise consequential actions. Screen control is not protected by the same tool-by-tool permission checks as ordinary file operations.

Organizations have a different set of controls

The Pro/Max removal should not be generalized to managed accounts. Team cloud sessions are enabled by default, while Enterprise cloud sessions are disabled by default. Owners have a separate cloud toggle; Enterprise enablement also requires appropriate role grants. Evaluate those settings rather than treating a larger personal subscription as an administration solution.

Two monitoring details matter. Endpoint detection tools cannot inspect the cloud runtime or the local execution VM. And Cowork’s network-egress policy does not cover web fetch, web search or MCP tools. Review those tool permissions separately; network-policy changes apply to new sessions.

Team and Enterprise can configure OpenTelemetry export for visibility. It can include prompts, tool parameters and file paths, so the collector becomes another destination for task information. Set its retention, access and redaction rules before enabling export; no events flow by default without administrator configuration.

What to do with your existing workflows

  1. Identify the requirement you actually have: cross-device continuation, access to one computer, offline data handling or organizational oversight. Use the comparison above to choose an execution path.

  2. For each recurring task, list its input location, tool dependencies and output destination. Review local-file schedules individually, and keep a manual fallback until the documented ambiguity is resolved.

  3. Before expanding usage, review connected folders and apps, output accuracy and plan consumption in Settings > Usage. Multi-step tasks consume more usage than quick questions; a subscription tier is not a guaranteed number of completed jobs.

Methodology: AI-assisted reporting and analysis based on Anthropic’s published documentation, rechecked on October 6, 2026. No authenticated account, migration or device workflow was tested. Account-level rollout and the conflicting local-schedule instructions remain unverified; no performance or security outcome is claimed.