Anthropic Just Gave You 3 Tools That Work While You're Gone.

Anthropic’s Dispatch plus “computer use” turns Claude into a genuinely remote, always-on work agent—letting tasks run while someone is away, not just generate text to read later. The core shift is practical: instead of producing summaries or drafts that still require human follow-through, Dispatch is positioned as a way to schedule work, persist context across sessions, and even operate desktop apps (including tools without APIs) so completed outputs land when the person returns.

The argument centers on two “primitives” that make this possible. First is cloud scheduled tasks: Claude can run recurring jobs on Anthropic’s infrastructure even when a laptop is off, with configurable network access, environment variables, and setup scripts. The schedule is meant for hourly or multi-hour recurrence rather than constant polling, and it can plug into existing MCP-based integrations (Linear, GitHub, Slack, OpenBrain, Google Drive, and more) without reconfiguring connectors. Examples include waking up to AI news already parsed and condensed, monitoring flight prices on a schedule and alerting when routes drop below a threshold, and handling recurring personal obligations like bill reminders.

Second is persistence through Dispatch. While it’s easy to describe as “persistent chat on your phone,” the more consequential detail is orchestration: a mobile conversation can spawn and manage multiple Claude work sessions in parallel on the desktop. Each session keeps its own context, file access, and connectors, with the phone acting as the command surface and the desktop acting as the execution surface. That parallelism is framed as the real productivity unlock—freeing people from the dead time at the desk while Claude tokenizes or works through multi-step tasks. A key constraint remains: the desktop must be running, and there are early limitations like per-subtask folder access approvals, no direct file attachments from the phone, and no direct output return to the phone. Workarounds include syncing outputs through Google Drive or Dropbox. Complex multi-app tasks are also described as roughly a 50% success rate in early testing, which is why Dispatch is labeled a research preview.

The most builder-relevant piece is “computer use” for apps without MCP servers. Even though MCP is treated as the “universal USB” for AI integrations, coverage will never be complete. Dispatch and computer use aim to bridge that gap by enabling keyboard-and-mouse automation in a remote browser/session, so Claude can handle legacy systems and bespoke workflows—like pulling data from outdated Jira instances, ERP screens, or hard-to-integrate accounting portals—then deliver the results into a spreadsheet workflow.

Beyond product mechanics, the transcript pushes a management mindset: delegate the “open loops” that sit in people’s heads (promises, deadlines, missing minutes, overdue scope updates) and decisions that get made with too little information. The best use cases are framed as work that compounds over time via scheduled research and OpenBrain-style knowledge accumulation. The closing warning is psychological as much as technical: humans will need to trust that work is happening while they’re away, learning to “walk away” rather than repeatedly checking whether the agent is truly doing the job.