We tested Claude Cowork’s computer use mode and its Dispatch functionality, which allows you to control your computer remotely from your smartphone. Three use cases and one observation: AI still achieves its goals, but not yet at the speed of a human.
Tasks that complete on their own, on your computer, without your help. This is the promise of Claude Cowork’s new computer use mode. It is now possible to control Anthropic’s AI agent from your phone using the Dispatch functionality. Sorting documents, generating reports, presentations, monitoring… Everything, or almost everything, becomes accessible in one click from your smartphone.
Computer use, a feature reserved for Macs
To understand Dispatch, you first need to understand the underlying architecture. Claude Cowork is based on two distinct layers. The first is direct connectors: native integrations with services like Slack, Google Calendar, Gmail or Google Drive, which allow the agent to act on these applications without ever touching the graphical interface. The second layer is the computer use itself: when no connector is available, Claude takes direct control of the machine. It controls the mouse, types on the keyboard, opens windows, navigates in a browser, interacts with any open application. It is this second layer that makes the difference with a simple conversational agent: the AI is no longer confined to the tools that we have explicitly connected to it.
Dispatch mode then allows you to control the session from your phone using the Claude application. Once the two sessions are linked, the user sends an instruction from their phone, and the agent executes it autonomously on the desktop. Claude decides the most efficient path himself: direct connector if available, otherwise control of the interface. It can chain several applications within the same task, for example extracting data from a spreadsheet, formatting it in a presentation, then sending the file by email. The only limitation is that computer use mode is only available on Mac, for the moment.
The JDN test
To test Claude Cowork’s capabilities in computer use mode, we submitted three different use cases to it, from our mobile. The first: a sectoral watch carried out on the web with a structured restitution in a Google Doc. The second: an accounting balance sheet generated in a Google Sheets from around ten invoices. The third: a pre-meeting preparation brief, combining agenda, emails and web searches.
To begin, you must link the Claude Desktop application with the Claude mobile application (iOS or Android). In Claude Desktop, click on the “Dispatch” button, then on “Start”. You must then authorize screen recording in the settings and grant additional access to Claude. Once this configuration is complete, the phone is connected to the computer and testing can begin.
As the prompts are intended to be typed on a smartphone, we deliberately use minimalist, very short prompts.
1. A report based on sector monitoring
Prompt: Keep an eye on the latest news on generative AI in business (last 5 days). Synthesizes the 5 key trends and creates a Google Doc entitled ‘AI Monitoring – (current date)’ with a summary structured by theme, sources and links.
Once the prompt is sent via the mobile application, Claude is discreet on screen. Operations that can be internalized without taking control of the computer are. Claude only takes control of the desktop again to write his report directly in Google Drive: the AI writes section by section and uses the Google Docs menus for formatting. The final result is correct, but the execution remains quite long (around 10 minutes).
2. An accounting balance sheet from an invoice file
Prompt: Opens the Invoices folder on my desktop. For each invoice, extract the supplier, the date, the amount excluding tax and tax included and the expense category. Create a Google Sheets ‘Balance sheet (months)’ with all this data, a total by category and an overall summary at the bottom.
For this use case, Claude further minimizes the use of the graphical interface and navigates primarily via internal commands. Surprising choice: while he could have queried the files via the CLI, he prefers to use the Finder to locate the invoices. Reading then takes place internally, nothing appears on the screen. Claude only takes control of the computer again to paste in the information from the accounting report that he himself produced. Using Sheets is clearly not comfortable for him: he struggles to navigate and paste his data properly. He first inserts a raw block, then attempts to create columns via the auto-table functionality, and then ends up resorting to macro scripts (Claude always comes back to code when he can!). It will manage to generate a fairly usable accounting table, but at the cost of 15 minutes of work. In this use case, a human would have been more efficient.
3. A brief before meeting
Prompt: Look at my next appointment in Google Calendar. Search the web for recent information about the participants and their company, consult the latest Gmail exchanges with them, and generate a Google Doc ‘Brief (meeting name)’ with: context, key points to discuss, history of exchanges and 3 questions to prepare.
Once the prompt is sent from the mobile, the Desktop version of Claude still remains quiet. The connection to Google being native in Cowork, the AI uses its own connectors directly and avoids going through the browser: it’s smart. Total time: 7 minutes. The brief is generated cleanly and responds perfectly to the instructions.
Should we adopt Claude Dispatch?
Claude achieves his goals in our three use cases, without exception. However, it is significantly slower than a human in an equal position. This is explained on the one hand by the numerous security measures deployed natively by Anthropic to supervise the agent. And on the other hand, by the relative slowness of the model, which must reason and plan each action before executing it to limit errors as much as possible. There is no doubt that the next versions of Claude will gain in intelligence and speed, and will erase these initial flaws.
However, should we adopt Dispatch on a daily basis? The answer depends on the context. For a user on the move, in transport or between two appointments, being able to launch a complex task from their phone and find it completed when arriving at the office is a real gain. On the other hand, at your workstation, it will almost always be faster to do the work yourself or to go directly through the Claude chatbot, which is much more responsive. The computer use mode nonetheless remains a convincing demonstration of what the personal agentic AI of tomorrow will be: an assistant capable of acting on your machine, not just responding to you.
