Claude Cowork realizes the fantasy that was held until now, without reality, by the majority of AI publishers on the market: having a universal agent capable of carrying out any task or almost any task.
The AI agent capable of doing your shopping as well as writing a financial report or sorting your emails is here. Marketed for several years by AI publishers as a distant promise, Anthropic is making it a reality. With Cowork, the San Francisco scale-up offers a universal agent that can be used by the majority of users, without advanced technical configuration. The JDN was able to test the tool.
A Swiss army knife AI agent on your PC
Surprised by the misuses of Claude Code by part of the developer community (writing reports, advanced online search, generating Excel files, etc.), the Anthropic teams designed an alternative to Claude Code to put the main features of Claude Code into the hands of non-technical users, with a simple graphical interface. Cowork takes the form of a conversation where the user can ask the agent to perform actions in the background.
The system then creates a virtual machine to execute all the actions inherent in resolving the final task. The agent reasons step by step and uses a multitude of tools and connectors to achieve its ends. It works on the same technical basis as Claude Code. Several actions can be launched in parallel to further optimize your time. The tool is currently available to subscribers of Claude’s Max offer (from 100 dollars per month) and Pro (20 dollars per month), only for Claude Desktop under macOS. It is an expensive investment that may be worth it for demanding users who plan to delegate many tasks to Claude Cowork.
An interface dedicated to the Cowork agent
Since the last update of Claude Desktop, Cowork arrives in the form of a third interface in Claude Desktop: there are classic chats, a GUI version of Claude Code, and now Cowork. Cowork was developed by the Anthropic teams as a true agentic experience. The main interface, the chat, allows you to send prompts as in a classic conversation with Claude. In the right column, several sections appear:
- Progress: which allows you to track where the agent is in relation to its initial plan
- Artifacts: which allows you to view at a glance the graphics, possible codes and visual elements generated by the AI during its journey
- Context: tools and files currently used by AI
- Suggested connectors: native or third-party connectors
By default, Claude Cowork can access the entire hard drive, but it is possible to partition the access directory to prevent the AI from acting or consulting files and folders not affected by your request. Finally, all work sessions take place on a virtual machine in the background, thus avoiding direct modification of your system.
The JDN test
To test Cowork’s capabilities, we put it through 3 tasks with different levels of complexity.
1. Sort your desk
We started with the classic and very simple case presented by Anthropic: sorting and tidying your desk. So we ask the AI to sort the files and folders on the desktop and remove unnecessary items. The agent begins by listing all the files and folders present, then asks us, via a dedicated form, which files are considered unnecessary. We tell it our preferences and the AI agent continues. The end result is perfect: a sorted desk without clutter.
2. Sort emails and suggest responses
Another use case, slightly more complex: sorting emails and generating a draft response. We asked Claude to consult our Gmail box and respond (without sending the response) to the last 10 emails received. Here again, Claude asked us about the style preferences to adopt. Once the criteria are sent, Claude starts using Chrome (via a widget in the browser) and navigates our inbox autonomously in the background. It is possible to see live what Claude is doing on your browser or to minimize the window to let him operate in the background. The end result is very clean, with adapted and perfectly calibrated responses for each email.
3. Create a report from a voice recording
More complex and more interesting, we ask Claude to create a complete report from a voice recorded meeting, an MP3 file. How will AI proceed without a transcription tool on hand? Claude begins by establishing a plan and surprises us. He understands that he must transcribe the audio file before he can use it to create his report. To do this, it cuts, in complete autonomy and without us asking it, our 56-minute file into 12 segments of 5 minutes and decides to transcribe them with Whisper, with the Tiny version to go quickly, locally from our machine. The approach is intelligent. Once the 12 transcriptions are completed, he combines them into a single file which he then uses to write his report. The report is then written using an Anthropic Word report writing skill.
The final report is of surprising quality and perfectly summarizes the comments made during the meeting. The report format is fully respected and the information is very readable and directly usable. The whole thing only took a few minutes and could have taken several hours of writing for a human.
A versatile agent aimed at non-specialists
Claude Cowork is a truly versatile AI agent that keeps its promises. Anthropic delivers a functional and accessible tool that will impress non-technical users. The “wow” effect is there: seeing the AI sort your desk, manage your emails or create a complete report from an audio recording without human intervention is still magic for many. The tool is aimed at the average worker looking to increase productivity on a daily basis, without having to manage complex configuration. Developers and technical profiles will find the same functionalities and results with Claude Code, which is more suitable.
The only real downside concerns data confidentiality. All personal information on your computer (files, emails, documents) is transmitted to Anthropic’s servers to allow the agent to function. This architecture may pose compliance and security issues depending on your industry. Companies subject to strict regulations (health, finance, legal) will probably have to abstain, unless Anthropic eventually offers a version with data processing locally or on private infrastructure.
