Anthropic Cowork feature creates 10GB VM bundle on macOS without warning

When AI Tools Silently Eat Your Storage: The Hidden Cost of Background Features

Imagine opening your Mac's storage settings and discovering that 10 gigabytes of disk space have vanished — consumed by a feature you never explicitly activated. That's exactly what macOS users recently experienced with Anthropic's Claude Code Cowork feature, which quietly spins up a virtual machine bundle that balloons to 10GB or more without a single warning dialog. For developers working on machines with 256GB or 512GB SSDs, that's between 2-4% of total storage silently claimed. The incident has sparked a broader conversation about software transparency, resource consent, and what users should expect from the AI tools they invite onto their machines.

What Actually Happened: The 10GB Surprise

Anthropic's Claude Code — a command-line AI coding assistant — introduced a feature called Cowork, designed to give the AI agent a sandboxed environment for executing code and managing files. Under the hood, this feature provisions a lightweight virtual machine on macOS, packaged as a .bundle directory that can swell to 10GB or more depending on usage. The VM includes a minimal Linux environment, dependencies, and workspace files necessary for the agent to operate autonomously.

The problem wasn't the feature itself — sandboxed execution environments are a legitimate and often preferred approach to AI agent safety. The problem was that users were never presented with a consent dialog, a storage estimate, or even a notification that a multi-gigabyte resource was being created on their system. Developers discovered the bloated bundles only after investigating unexpected storage warnings or running disk analysis tools like DaisyDisk or du -sh commands in their terminal.

Community discussions revealed that many users had no idea the VM existed until their machines started running low on space. Some reported finding multiple bundles from separate sessions, compounding the storage impact to 20-30GB in extreme cases. For anyone who has ever frantically cleared space before an important deployment or client demo, the frustration is immediately relatable.

Why Silent Resource Consumption Is a Growing Problem

This incident isn't isolated. As AI-powered developer tools become more sophisticated, they increasingly require substantial local resources — model caches, vector databases, container images, and execution environments. GitHub Copilot's local indexing, Docker Desktop's VM overhead, and various AI coding assistants all consume meaningful disk space. But the difference between acceptable and unacceptable resource usage almost always comes down to one factor: informed consent.

Users expect a download manager to tell them a file is 2GB before they click "download." They expect a game on Steam to display its storage requirements before installation. Yet many developer tools — especially those in rapid iteration cycles — skip these courtesies in favor of seamless experiences. The assumption is that developers are technical enough to manage their own systems. But "technical enough" doesn't mean "psychic."

The most dangerous assumption in software design is that your users will discover and manage the side effects of features they never knew existed. Transparency isn't a nice-to-have — it's the baseline of trust between a tool and the professional who depends on it.

The Real Cost: More Than Just Gigabytes

Ten gigabytes might sound trivial in an era of terabyte drives, but the real-world impact extends far beyond raw storage numbers. For developers and small business operators, the consequences cascade in several ways:

• Time Machine and backup bloat: macOS Time Machine backs up these VM bundles by default, meaning the 10GB multiplies across every hourly snapshot — potentially consuming 100GB+ of backup drive space within days
• iCloud sync conflicts: Users with Desktop or Documents folder sync enabled may inadvertently push gigabytes of VM data to iCloud, burning through their cloud storage quota
• CI/CD pipeline impact: Developers who mirror their local environment to build servers may unknowingly transfer unnecessary VM artifacts
• SSD wear: Repeated creation and deletion of large files accelerates write wear on solid-state drives, particularly on older machines with less durable NAND
• Performance degradation: On machines with limited RAM, the VM's background processes compete for memory, causing swap usage to spike and overall system responsiveness to drop

For freelancers and small business owners running their operations from a single MacBook — managing client work, invoicing, project management, and now AI-assisted development — every gigabyte matters. When your machine is also your business hub, unexpected resource consumption directly impacts productivity.

What Good Software Transparency Looks Like

The fix for this class of problem isn't complicated. It's a design philosophy that prioritizes the user's awareness and control over convenience and seamlessness. Software that respects its users follows a predictable pattern:

1. Disclosure before action: Before creating resources that exceed a reasonable threshold (say, 500MB), present a clear dialog explaining what will be created, how much space it will consume, and where it will be stored
2. Opt-in over opt-out: Resource-heavy features should require explicit activation rather than running by default. A single toggle in settings is all it takes
3. Visibility in context: Provide a dashboard or status indicator showing current resource usage. If your tool is consuming 10GB, surface that information where users naturally look
4. Clean uninstall paths: When a user disables a feature or uninstalls the tool, all associated resources should be removed — or at minimum, the user should be prompted about leftover data

This philosophy extends beyond AI tools to any business software. When platforms handle sensitive operations — financial data, customer records, automated workflows — transparency about what's happening under the hood isn't just good UX. It's an ethical obligation. At Mewayz, this principle drives how we architect our 207-module business OS: every automation, every data sync, every resource allocation is surfaced in the activity dashboard so users maintain full visibility over what their platform is doing on their behalf. No silent processes, no hidden storage consumption, no surprises.

Lessons for the AI Tool Ecosystem

The broader AI developer tools market is moving at breakneck speed. New features ship weekly, capabilities expand monthly, and the competitive pressure to deliver "magical" experiences often overshadows the mundane but critical work of resource management and user communication. Anthropic is far from the only company navigating this tension — but high-profile incidents like this serve as useful inflection points for the entire industry.

The developers most affected by silent resource consumption are, ironically, the exact users these tools are built for: professionals who rely on their machines to run multiple demanding applications simultaneously. A web developer might be running a local dev server, a database instance, a design tool, a communication app, and an AI coding assistant all at once. When any one of those tools silently claims 10GB without warning, it disrupts the careful balance that professional workflows depend on.

To Anthropic's credit, the company has been responsive to community feedback in the past, and this issue is likely to be addressed with clearer disclosure in future releases. But the pattern — ship fast, discover side effects through user complaints, patch after the fact — is one the industry needs to outgrow. Proactive transparency should be part of the development process, not a reactive fix.

Protecting Your Machine: Practical Steps

Whether you've been affected by this specific issue or simply want to guard against similar surprises from any tool in your stack, there are concrete steps you can take today:

• Audit your disk regularly: Use tools like DaisyDisk, GrandPerspective, or the built-in macOS Storage Management to identify unexpected space consumption at least monthly
• Exclude VM bundles from backups: Add known VM directories (like ~/.claude/ or ~/Library/ subdirectories) to your Time Machine exclusion list if you don't need them backed up
• Monitor with terminal commands: A simple du -sh ~/Library/* | sort -hr | head -20 command reveals the largest consumers in your Library folder
• Use cloud-first tools where possible: Business platforms that run primarily in the cloud — like Mewayz's browser-based workspace for CRM, invoicing, HR, and project management — keep your local machine lean while still delivering full functionality across all 207 modules
• Set storage alerts: Configure macOS notifications to warn you when available storage drops below 15-20% so you catch unexpected consumption early

For teams and businesses, establishing a standard operating environment policy that documents which tools are approved and what resources they typically consume can prevent these surprises from cascading across an entire organization. A single developer discovering a 10GB surprise is an inconvenience; fifty employees across a company each losing 10GB is a support ticket avalanche.

The Bigger Picture: Trust as a Feature

At its core, the 10GB VM incident is a story about trust. Every piece of software you install is making an implicit promise: "I will do what you expect, and nothing more, without your knowledge." When that promise is broken — even unintentionally, even for technically sound reasons — it erodes the relationship between tool and user. And in a market where developers and business operators have dozens of alternatives for any given tool category, trust is the feature that determines long-term adoption.

The most successful software platforms of the next decade won't just be the most powerful or the most AI-enhanced. They'll be the ones that treat their users as partners deserving of full transparency. Whether you're building a virtual machine sandbox, an automated invoicing workflow, or a customer relationship pipeline, the principle is the same: tell people what you're doing, give them control over it, and never assume silence equals consent. That's not just good software design — it's good business.

Frequently Asked Questions

What is the Anthropic Cowork feature and why does it use so much space?

The Cowork feature in Claude for desktop is designed to provide a fully containerized development environment. This requires creating a large virtual machine (VM) bundle, which includes an entire operating system and necessary tools. These VM images are inherently large, often starting at several gigabytes, which explains the significant 10GB+ storage consumption reported by users.

How can I check if Cowork is using storage on my Mac and how do I disable it?

You can check your storage in About This Mac > Storage. To disable it, open the Claude desktop app, go to Settings, find the Cowork feature, and toggle it off. This should prevent the VM from running and allow the system to eventually clean up the allocated space, though manually deleting the bundle might be necessary.

Are there lightweight alternatives for AI-powered coding assistance?

Yes, many coding assistants operate without requiring a full local VM. For instance, cloud-based tools or extensions that integrate directly with your IDE typically have a minimal footprint. Services like Mewayz, which offers access to 207 AI modules for $19/mo, provide powerful coding assistance without consuming significant local storage resources.

What should developers consider before using AI tools with local execution?

Always check the system requirements and documentation for any background processes or local installations. Be mindful of your available storage, especially on devices with smaller SSDs. Weigh the benefits of local processing (like offline use) against the resource cost. Exploring alternatives, such as Mewayz's extensive module library, can offer robust features without the hidden storage tax.