AI startup Anthropic recently quietly announced a new feature called "Dispatch" on the X platform, allowing users to directly pick up their phones and remotely control their awake Macs at home or in the office with a single sentence, so that Claude on the Mac can automatically perform various desktop tasks for them.
Breaking screen limitations, a single sentence can activate remote AI workforce.
Imagine this scenario: You're crammed onto the subway during your commute, and suddenly remember you haven't organized yesterday's meeting notes. Simply take out your phone and say to the Claude App, "Organize yesterday's meeting recording into Notion." A few minutes later, Claude will report that the task is complete.
We're shipping a new feature in Claude Cowork as a research preview that I'm excited about: Dispatch!
One persistent conversation with Claude that runs on your computer. Message it from your phone. Come back to finished work.
To try it out, download Claude Desktop, then pair… pic.twitter.com/r6OH46Ll89
— Felix Rieseberg (@felixrieseberg) March 17, 2026
This isn't science fiction; it's a scenario that Dispatch is actually implementing.
Dispatch's core concept is completely different from the traditional "remote desktop." A remote desktop still requires you to manually move the mouse and type on the keyboard; you're just viewing it on a different screen. Dispatch, on the other hand, turns your phone into a "pure command execution device," and Claude on your Mac is the real "executor." It automatically clicks, types, and performs cross-software operations, so you don't need to look at the screen at all.
The setup process is surprisingly simple: just update the Claude desktop version (Cowork mode) on your Mac, click the newly added Dispatch option, generate a QR code, and then scan and pair it using the Claude mobile app. The only prerequisite is that your Mac must remain awake (not asleep) and the Claude application must remain open.
The actual success rate is only 50%, but why is it so significant?
Currently, Dispatch is still marked as a "Research Preview" and is only available to users of the highest-tier Max subscription plan. According to some media tests, the performance of this feature is almost "mixed," with an overall success rate of only about 50%.
• What can be done:Precisely identify images containing specific keywords in screenshots, manipulate Notion (list notes or add URLs), and read and summarize recently received emails.
• It is impossible to do:It may fail to open certain applications (such as Mac Shortcuts), send messages across applications (such as sending a screenshot to a colleague via iMessage), and encounter failures when loading Safari browser tabs or services that require third-party authorization.
In traditional software evaluation, a 50% success rate is obviously failing, but in the context of AI Agent development, it is a highly significant milestone.
Looking back at the evolution of AI's "operational radius": In 2023, AI could only produce text and code in dialog boxes; in 2024, AI began to generate web page elements (such as artifacts); in early 2025, Anthropic launched the Cowork feature, which could take over the mouse and keyboard, allowing AI to enter the entire operating system desktop. Now, Dispatch has successfully crossed the boundary of "physical devices," enabling asynchronous proxy operations across devices.
"Local First" Architecture: Anthropic's Smartest Security Defense
While letting an unseen AI remotely take over your personal computer sounds terrifying, Anthropic made a crucial and clever architectural decision to address this trust issue: completely locking the execution end locally.
Unlike many remote AI solutions that upload screenshots to cloud servers for analysis, Dispatch takes a different approach: your phone sends a natural language command → the command is transmitted to your Mac → Claude on your Mac executes the operation in a local sandbox environment. This means that your private emails, work files, and screen remain on your Mac from start to finish, significantly reducing the risk of privacy leaks.
Analysis of viewpoints
Anthropic has been making frequent and highly strategic moves recently. From "Remote Control," launched at the end of February for engineers to remotely control terminals, to "Dispatch," for general users to remotely control graphical user interfaces (GUIs), Anthropic is attempting to seize the dominant position of "cross-device general-purpose AI agent" at a time when OpenAI is choosing to scale back its efforts and focus on enterprise and coding capabilities.
Although Dispatch currently has a success rate of only 50%, this number is bound to rise rapidly within a few months as the model is iterated and updated. What we should really be thinking about is: when everyone has a Mac at home that's on 24/7, running a readily available, unpaid, and uncomplaining "digital clone," how drastically will our definition of "work" and "labor force" be transformed?
