Anthropic Just Moved Claude Cowork to the Cloud. That's a Bigger Deal Than You Think.

Your laptop is closed. You're asleep. And somewhere in an Anthropic data center, your AI agent is still working — running the task you scheduled at 3 PM, browsing the web, processing files, and waiting to ping your phone when it's done.

That's not a hypothetical anymore. As of this week, it's what Claude Cowork does.

Anthropic just moved Claude Cowork to the cloud by default. Previously, it was a desktop-only feature — you needed the Claude app running on your Mac or PC. Now, agent sessions live on Anthropic's servers. They persist across devices. They run in the background. Scheduled tasks fire even when every single one of your devices is offline.

This is a bigger deal than the "now on mobile" headline suggests. Anthropic just made a clear architectural bet: the future of AI agents is in the cloud, not on your device.

Let's talk about why that matters — and why they might be half right.

What Actually Changed

Claude Cowork, for the uninitiated, is Anthropic's agentic computing feature. Instead of chatting with Claude turn by turn, you hand it a task — "research these three competitors and build a comparison doc," "monitor this GitHub repo and summarize new issues daily," "clean up my calendar and draft responses to pending emails" — and Claude goes off and works on it autonomously. It browses, reads files, runs code, and reports back.

Until now, that all happened locally. Your desktop Claude app was the execution environment. Close the app, the agent stops. Simple, but limiting.

The update changes three specific things:

1. Cloud-first execution. Cowork sessions run on Anthropic's infrastructure by default. You can still opt for local processing on the desktop app, but cloud is the new default — which tells you where they're steering the product.
2. Scheduled tasks without devices. You can schedule a task to run at 6 AM, and it will — even if your laptop is off, your phone is on a nightstand, and you're dead to the world. The cloud session doesn't need you. It just needs the schedule.
3. Cross-device continuity. Start a task on your Mac, check on it from your phone, review the results on the web. The session lives in the cloud, not on any one machine. Anthropic is also adding push notifications — when Claude finishes something or needs your approval, it pings your phone.

This is rolling out to Max subscribers first (their highest tier), with other plans "in the coming weeks." Anthropic also extended doubled usage limits through August 5th — probably to get people hooked on the cloud workflow before the meter returns to normal.

The Cloud Agent Thesis

Here's what's really interesting: Anthropic didn't just add a mobile app. They made a deliberate architectural choice about where compute happens. And that choice says a lot about where they think this is going.

The cloud-agent thesis is simple: AI agents are most useful when they're always available. Not just when you're at your desk. Not just when your laptop is open. Not just when you remember to launch the app. An agent that can fire at 6 AM, work for two hours, and have results waiting when you grab your phone over coffee is fundamentally more useful than one that needs you present to babysit it.

Think about it from the user's perspective. The whole pitch of agentic AI is that it saves you time. It does the tedious stuff while you do other things. But if "while you do other things" means "while your laptop is open and the app is running," that's not really autonomous. That's supervised execution with extra steps.

Cloud execution breaks that constraint. Your agent becomes a background worker. It has a schedule. It runs independently. It notifies you when it's done. This is closer to hiring a virtual assistant than installing a tool.

This is probably correct for a lot of use cases. Scheduled research, monitoring tasks, batch processing, report generation — these are things that benefit enormously from being decoupled from your physical presence.

But here's where it gets complicated.

The Trade-off Nobody's Talking About

Cloud-first agents solve the availability problem. They create a different one.

When your agent runs locally, it has access to your files, your terminal, your applications, your local network. It operates inside your environment. The data stays on your machine. You can unplug the internet and it still works. You can audit exactly what it did because it did it on hardware you control.

When your agent runs in the cloud, it operates inside Anthropic's environment. It can browse the web and access files you upload, but it can't open your local spreadsheet. It can't run a script on your machine. It can't read your calendar through the native API. It can't execute terminal commands on your system. The "full experience" — Anthropic's own words — is still desktop-only, including local file access.

So Anthropic is trying to have it both ways: cloud for always-on convenience, desktop for actual deep work. And that split creates a weird user experience. Your scheduled background research runs in the cloud. But the task that needs your local repo, your local files, your local tools? Back to desktop. Back to your laptop being open. Back to babysitting.

This isn't a knock on Anthropic. It's a fundamental tension in agent architecture that nobody has solved. You can have always-on convenience (cloud) or deep system access (local). Picking one means compromising on the other.

Why the Local-First Crowd Isn't Wrong

There's a reason products like CopperRiver, Raycast AI, and Apple's on-device intelligence bet on local execution. Several reasons, actually.

Privacy. When your agent runs in the cloud, your data — your files, your browsing, your schedule — leaves your machine and lands on someone else's servers. Anthropic is a well-run company. But "trust us with everything" is not an architecture; it's a prayer. Regulated industries (healthcare, finance, legal), privacy-conscious users, and anyone who's been paying attention to the last decade of data breaches has legitimate reasons to want their data to stay on their hardware.

Latency. Cloud agents have network round-trips for every action. Every file read, every command, every browser action goes through the network. Local agents don't. For interactive tasks — the kind where you're iterating quickly, watching the agent work, adjusting mid-task — the difference is the gap between "feels instant" and "feels like loading."

Control. A local agent can't be rate-limited by a provider. It can't have its plan downgraded because a model is overloaded. It can't be shut off because your subscription lapsed. It can't change its behavior because the provider updated the model. It's yours, running software you chose, doing what you told it to.

Offline capability. Not everyone has perfect internet. Not every task needs the internet. An agent that can work on your local files without a connection is more resilient than one that goes dark when the Wi-Fi drops.

Cost. Cloud agents bill per token, per action, per API call. Local agents run on hardware you already own. If you're doing serious agent work — dozens of tasks a day, long-running sessions — the economics of local start looking very different from the economics of a $200/month cloud subscription.

The local-first thesis is: the most sensitive, most interactive, most system-integrated tasks should run on your machine, where you have control, speed, privacy, and predictable costs. The cloud is for things that genuinely benefit from being always-on and web-connected.

The Honest Assessment

Here's where I land: Anthropic's cloud shift is the right call for their product. Claude is a cloud-native tool. Their model runs in the cloud. Their users are already sending data to Anthropic's servers with every message. Moving agent execution there too is consistent and convenient. For Claude's target user — someone who wants a smart assistant and doesn't deeply care where the compute happens — this is strictly better than the old model.

But it's not the only valid architecture. And it's not the right architecture for everyone.

The users who care about data sovereignty, who want their agent to have deep system access, who want to run open-source models locally instead of paying per-token for a proprietary API — those users aren't served by cloud-first agents. They're served by local-first tools that bring the model to the data, not the other way around.

The interesting question isn't "cloud or local?" — it's "what runs where?" The answer that makes sense is hybrid: cloud for scheduled, always-on, web-dependent tasks. Local for sensitive, interactive, system-integrated work. Neither side wins entirely. The best products will be the ones that let users choose per-task, not per-product.

Anthropic's update is a step toward one end of that spectrum. It's a good step for them. But the other end of the spectrum — powerful, private, local-first agents — is just as important and just as underserved.

What This Means for the Market

Anthropic moving Cowork to the cloud validates the agent computing thesis hard. They're not building a chatbot anymore. They're building infrastructure for autonomous digital work. The fact that they're willing to run compute on their servers while you sleep — on their dime, within your subscription — tells you how seriously they take this shift.

OpenAI is presumably watching. Google has the infrastructure to do the same thing tomorrow. The cloud-agent wars are starting, and they're going to be fierce.

But the move also opens space. If cloud agents are great for scheduled tasks but bad for local system access, then the product that does local well — that gives you an agent with real terminal access, real file system control, real browser automation, running models you chose, on your hardware — that product serves a different market. A market that's growing as people get more sophisticated about what they want from AI tools and more skeptical of sending everything to the cloud.

The agent war isn't cloud vs. local. It's about who gives users the most capability with the least compromise. Anthropic just made their move. The local-first crowd's response is going to be interesting to watch.

One thing is clear: the era of agents that need you to hold their hand is ending. Whether they grow up in the cloud or on your desktop — that's the real question.

---

If you're on the local-first side of this debate — or you're just tired of sending every file to the cloud — CopperRiver runs AI agents directly on your Mac. Real terminal access, real file system control, real browser automation, with open-source models like GLM, DeepSeek, and Qwen. Your data stays on your machine. Plans start at $9/mo.