A Quote from Felix Rieseberg
Anthropic moves Cowork's VM tool execution to cloud sandboxes, cutting local disk and battery use while the desktop app still handles file access.
Key Takeaways
- Anthropic moved Cowork’s virtual machine execution to cloud sandboxes, so both model inference and tool calls now run remotely instead of on the user’s computer.
- The desktop app still handles local file-access requests, preserving a permission boundary between cloud execution and the user’s device.
- This reduces local disk use, battery drain, and performance overhead, while allowing tasks to continue even if the laptop is closed or the user switches devices.
- Each Cowork session runs in an isolated cloud sandbox, meaning working state is not shared with other sessions.
- The update shifts Cowork’s core execution burden to the cloud while keeping safety, capability, and local file control as key design constraints.
Meta description: Anthropic now moves Cowork’s VM execution to cloud sandboxes, reducing local disk and battery use while keeping file access in the desktop app.
Tags: claude-cowork, anthropic, claude, generative-ai, general-agents
Anthropic has changed how Cowork carries out agent work. In the current setup, both model inference — the step where the AI produces its next response — and the virtual machine that performs tool calls run in the cloud. The desktop app still steps in when the cloud environment needs to touch a file on the user’s device. This update is aimed at people who use Cowork with Claude and want tasks to continue without draining a laptop’s disk, battery, and performance headroom.
What Changed in Cowork
Earlier Cowork split the workload in two places. The AI’s inference happened remotely, but the actions requested by the agent — known as tool calls — ran inside an Anthropic-provided VM installed on the user’s computer. That VM was not a casual add-on. Anthropic’s Felix Rieseberg explained that the company added it for capability, safety, and security reasons, limiting the environment so that only data explicitly added to the session was available inside it.
Now the execution environment has moved upward. The VM itself runs in the cloud, not as a local machine on the user’s computer. Cowork sessions are separated into individual sandboxes, or isolated execution environments, meaning the working state of one session is not shared with another. If the cloud VM needs to access something on the device, such as a local file, the desktop app remains responsible for carrying out that file access tool call.
Why the Old Local VM Was Useful — and Painful
The old design solved a real problem. Running tool calls inside a controlled VM let Cowork do more than a plain chat window could, while keeping the agent’s actions contained. For users who added only certain files or data to a session, the local sandbox created a boundary: the agent could work inside that boundary without wandering through the rest of the machine.
Still, local execution had costs. A VM needs disk space, memory, and processing power. On laptops, that can slow other apps, shorten battery life, and make the experience feel heavier. It also created a mobility problem. If the task depended on a VM running on the user’s laptop, closing the laptop could stop the work. The user had to keep the device available, even though the model itself was already being served from the cloud.
What This Means for Claude and Cowork Users
The new setup shifts the burden. Users who were frustrated by battery drain, disk use, or performance overhead should see the local machine become less central to Cowork’s operation. The same capability argument remains important: tool calls still run in an isolated environment, but that isolation is now provided in the cloud rather than through a VM shipped to the user’s computer.
File access is the key bridge. The cloud VM can ask for a file, but the desktop app is the part that handles the request on the device. That keeps a local permission boundary in place while moving the heavier execution away from the laptop. For people who use Cowork from a phone, the change is more than a convenience. It suggests that work does not need to depend on a local VM sitting on a particular machine.
What to Watch Next
The next question is how users experience the balance between cloud execution and local file control. If Cowork is used mainly for tasks that require many local files, the desktop app will remain an important permission checkpoint. If it is used for work that starts on a laptop and continues on a phone, the cloud sandbox could become the default mental model: a session that keeps running while the device changes.
Anthropic has also set an expectation about isolation. The company says each session gets its own sandbox and does not share state with other sessions. That matters for users who run multiple Cowork tasks at once. It also gives people a clearer way to understand risk: the boundary is no longer just the local VM on their machine, but a separate cloud execution session that still needs local file access when the task asks for it.
Comments (0)