For most of the past few years, AI assistants have been remarkably good at one thing: talking. Ask them a question, get an answer. Paste in some code, get it fixed. Upload a document, get a summary. The interaction always ended the same way — with words on a screen that you, the human, then had to go and do something with.
That changed on March 23, 2026.
Anthropic announced that Claude can now take direct control of your computer desktop — pointing, clicking, navigating menus, opening applications, filling spreadsheets, browsing the web, and completing multi-step workflows without you lifting a finger. It’s not a browser extension or a narrow automation script. It’s Claude, the same AI you’ve been chatting with, now able to operate your Mac the way a human assistant would if they sat down at your keyboard.
The feature, currently in research preview for Pro and Max subscribers on macOS, is available through Claude Code and Claude Cowork. And while Anthropic is careful to call it early-stage, the implications are anything but small. This is the moment AI stops being a tool you use and starts being a coworker you delegate to.
What Claude Computer Use Actually Does
The simplest way to understand Claude Computer Use is through Anthropic’s own demo. A user is running late for a meeting. From their phone, they message Claude: export the pitch deck as a PDF and attach it to the meeting invite. No further instructions. No hand-holding. Claude opens the file, exports it, attaches it to the calendar event, and reports back — done.
That’s the promise. And according to reporting from CNBC, it extends well beyond tidy demo scenarios. Claude can open apps, navigate a web browser, fill in spreadsheets, batch-resize photos, send files from your computer to your phone, and handle complex multi-app workflows that would otherwise require you to context-switch a dozen times.
The system works through a layered approach. Claude first looks for direct Connectors — purpose-built integrations with services like Google Workspace and Slack. If a Connector exists, it uses that. If not, it falls back to full screen control: taking screenshots, interpreting what it sees, reasoning about what to do, and then acting. As Ars Technica describes it, Claude can “scroll, click to open, and explore as needed” on the machine itself.
Central to the experience is Dispatch, a feature Anthropic released the week before the computer use announcement via Claude Cowork. Dispatch enables a continuous conversation with Claude across your phone and desktop. You assign a task — say, “research these five competitors and build me a comparison spreadsheet” — and you walk away. Claude keeps working on your desktop while you’re at lunch. When you come back, it’s done. The only requirement: your computer stays powered on.
This is what Anthropic means when it talks about an “AI coworker.” Not a chatbot. Not an autocomplete. A thing that does the work.
How We Got Here: A Very Fast Timeline
The speed at which this capability has materialized is genuinely striking. Anthropic shipped roughly 12 major features in about 12 weeks leading up to the March 23 announcement — an update cadence of almost every two weeks. Claude Code alone iterated from version 2.1.63 to 2.1.76 during this stretch.
Here’s how the run-up looked:
- January 2026: Claude Cowork launches — a more casual, non-developer version of Claude Code designed to bring agentic workflows to everyday users
- February 5: Claude Opus 4.6 releases, topping benchmarks in agentic coding, computer use, and tool use
- February 12: Anthropic closes a $30 billion Series G at a $380 billion valuation — one of the largest private fundraises in tech history
- February 17: Claude Sonnet 4.6 launches with enhanced computer use capabilities and a 1-million-token context window in beta, scoring 72.7% on OSWorld (the industry benchmark for computer use tasks)
- February (late): Remote Control ships as a research preview — letting you control your coding session from your phone after you’ve closed your laptop
- March 2: Claude Memory goes free for all users
- March 11: Excel and PowerPoint integrations arrive
- March 17: Persistent agent threads roll out for Pro and Max users
- March 23: Full computer use research preview goes live
The conceptual work started even earlier. According to Tech Insider, Anthropic’s research labs began experimenting with models that could interpret screenshots and take actions in late 2024. The initial beta was limited, but it proved the concept: large language models could go beyond text generation to become genuine computer operators.
Under the Hood: The Technical Architecture
Claude Computer Use doesn’t rely on APIs or pre-built hooks into applications. It works visually. Claude takes a screenshot, interprets what’s on the screen, decides what action to take, and executes it — then repeats the loop. This is why it works across virtually any application without special setup, and also why it’s slower and more error-prone than Connector-based integrations. Anthropic acknowledges this directly, noting that computer use “takes much longer and is more error-prone” than API-based approaches.
The model powering this on the consumer side is Claude Sonnet 4.6, which launched February 17 with significant upgrades to computer use and agent planning. For power users in Claude Code, the default model has been upgraded to Opus 4.6, which offers substantially deeper reasoning capabilities.
Claude Code version 2.1.76 — the developer-facing product — layers several capabilities on top of the base computer use feature:
/loopscheduled tasks: A cron-style scheduler built into Claude Code. You describe a task and an interval in plain language —/loop 15m check if PRs have new comments or CI failures— and Claude runs it in the background on a recurring basis. Tasks max out at 50 concurrent and expire after three days to prevent runaway loops.- Background Agents with Worktree isolation: Claude can run parallel sub-tasks in independent copies of your codebase so background work doesn’t interfere with what you’re actively doing.
- Remote Control: Start Claude Code in your terminal, then continue issuing instructions from your phone or the
claude.ai/codeweb interface after you’ve physically left your desk. Crucially, your code never leaves your local machine — only chat messages pass through an encrypted channel. - Voice Mode: A push-to-talk system (hold spacebar) supporting 20 languages, useful for dictating complex instructions when your hands are occupied.
Underlying all of this are meaningful model infrastructure upgrades: a 1-million-token context window now available on Max, Team, and Enterprise plans; 128K max output with Opus 4.6; and automatic context compaction to maintain coherence in ultra-long sessions.
The Competitive Landscape: Why Now
Claude Computer Use didn’t arrive in a vacuum. It arrived in the wake of OpenClaw — a third-party app that went viral in early 2026 by letting users message an AI through WhatsApp or Telegram to complete tasks on their local device. Nvidia CEO Jensen Huang called it on CNBC “definitely the next ChatGPT.” OpenAI subsequently hired OpenClaw’s creator, Peter Steinberger, specifically “to drive the next generation of personal agents.”
That single viral moment triggered a sprint across the entire industry. Within weeks:
- Perplexity launched Personal Computer, its own desktop agent
- Manus shipped My Computer
- Nvidia announced NemoClaw, an enterprise-grade OpenClaw competitor
- OpenAI has been expanding its Operator agent, which focuses on browser-based task completion
- Google’s Project Mariner (DeepMind) offers browser-level agent capabilities with the advantage of deep Chrome integration
- Microsoft’s Copilot Vision focuses on Office 365 workflows
Ars Technica noted that Anthropic is “joining the increasingly crowded field of companies with AI agents that can take direct control of your local computer desktop” — but the framing undersells Claude’s differentiation. Unlike browser-only competitors, Claude Computer Use operates at the full desktop level, across any application. And unlike Microsoft’s approach, which is tightly coupled to the Office ecosystem, Claude runs across Amazon Bedrock, Google Vertex AI, and Microsoft Foundry — making it platform-agnostic by design.
The 1-million-token context window is also a meaningful technical edge. Most competitors top out at 128K–256K tokens. Being able to ingest an entire codebase, months of emails, or a full legal document set in a single session without chunking or retrieval pipelines is a significant practical advantage for complex, long-horizon tasks.
The Safety Question No One Can Avoid
Giving an AI direct control over your computer is not a trivial decision, and Anthropic has been notably candid about the risks. The company explicitly warns that computer use “is still early compared to Claude’s ability to code or interact with text,” that Claude “can make mistakes,” and that its safeguards “aren’t perfect” and “aren’t absolute.”
As Ars Technica reports, the specific safeguards Anthropic has built in include:
- Prompt injection scanning to prevent malicious websites or apps from hijacking Claude’s actions mid-task
- Permission requests before Claude accesses any new app
- A list of “off-limits” categories blocked by default — investment and trading platforms, cryptocurrency apps
- Model-level training to avoid “risky operations” such as moving money, modifying files irreversibly, scraping facial images, or inputting sensitive data without explicit instruction
But there’s a hard truth buried in the fine print: when computer use is active, Claude can see everything on your screen — “personal data, sensitive documents, or private information,” in Anthropic’s own words. The company’s recommendation is to start with trusted apps and avoid sensitive data during this research preview phase.
This isn’t just a technical caveat. It’s a new category of trust relationship between user and AI. Cybersecurity researchers have long warned that the more autonomous access an agent has, the larger the potential attack surface. Every permission you grant Claude is, hypothetically, a permission an adversary could exploit if something goes wrong upstream.
Anthropic’s response to this concern is largely institutional: Constitutional AI as a foundational alignment methodology, the research-preview rollout strategy as a deliberate pacing mechanism, and a track record of publishing safety evaluations. Whether that’s sufficient for enterprise contexts handling sensitive data is a question every organization will have to answer for itself.
Enterprise Adoption: The Numbers Behind the Hype
While computer use is still in research preview, Claude’s broader agentic capabilities have already found substantial enterprise traction. According to Tech Insider’s deep-dive, major organizations including Rakuten, CRED, TELUS, Zapier, and Figma have deployed multi-agent coordination systems built on Claude. Figma integrated Claude Code directly into its design workflow. Box CEO Aaron Levie publicly stated that Sonnet 4.6 delivered “a 15% jump in performance and accuracy” in Box’s internal evaluations.
The financial context matters here. Anthropic’s annualized revenue run-rate reached $14 billion as of February 2026 — more than tenfold growth over three years. Claude Code alone surpassed $2.5 billion in annualized run-rate revenue, with that number doubling since the start of 2026. Weekly active users of Claude Code doubled year-over-year. Ramp’s spending data shows that one in five businesses now subscribes to Anthropic’s services, up from one in 25 a year ago.
These are not the numbers of a company selling a novelty. They’re the numbers of a company whose product has become operational infrastructure.
And then there’s the government angle. Claude is reportedly the only AI model in use for U.S. classified missions, through a Palantir partnership established in February 2026. A “Claude Gov” variant has been operational across multiple U.S. national security agencies since mid-2025. For enterprise buyers concerned about reliability and security track record, that credential carries weight.
What This Means for How We Work
The honest answer is: we don’t fully know yet.
What we can say is that computer use AI is explicitly designed to absorb the repetitive, screen-based layer of knowledge work — the part of the job that involves moving a file from one place to another, updating a spreadsheet with data from three different tabs, or sending a document that’s sitting on your desktop to someone who needs it right now while you’re stuck in traffic.
Anthropic CFO Krishna Rao framed it plainly: “Whether it is entrepreneurs, startups, or the world’s largest enterprises, the message from our customers is the same: Claude is increasingly becoming critical to how businesses work.” That’s a revenue-motivated statement, but it reflects something real about where enterprise AI adoption has landed.
Wharton professor and AI researcher Dr. Ethan Mollick has described this moment as the “third wave” of AI adoption: first chatbots, then coding assistants, now autonomous agents that can actually execute knowledge work. Tech Insider’s analysis echoes this framing — noting that the shift is less about replacing workers and more about eliminating the friction between deciding what to do and actually doing it.
The near-term reality is probably more mundane than the hype suggests and more significant than the skeptics allow. Dispatch and computer use won’t replace a skilled analyst — but they might free that analyst from the two hours a day they spend doing things that don’t require their expertise. At scale, across millions of knowledge workers, that’s not a small thing.
Availability and What Comes Next
Right now, Claude Computer Use is available to Claude Pro ($20/month) and Claude Max subscribers on macOS, through Claude Code and Claude Cowork. It is a research preview, which means Anthropic is actively gathering feedback, the feature will not always work perfectly, and complex tasks may require a second attempt.
Engadget confirmed that the feature also supports use with Dispatch — the cross-device conversation thread that lets you assign tasks from your phone and track progress across sessions. Windows support has not yet been announced with a firm date but is implied as a future target.
The /loop scheduled task feature and Background Agents are available on all paid Claude plans. The 1-million-token context window requires Max, Team, or Enterprise.
What to watch in the coming months: the transition out of research preview into a production-grade release; Windows availability; and how enterprise deployments handle the security and compliance questions that inevitably arise when AI has screen-level access to internal systems.
The Bigger Picture
There’s a version of this story that’s purely about competitive positioning — Anthropic shipping a feature to keep pace with OpenAI, Google, and the viral phenomenon of OpenClaw. That version is true but incomplete.
The more interesting version is about a fundamental change in what AI is. For the past several years, the implicit model has been: AI as a very smart search engine, or a very fast junior analyst. You ask it things. It tells you things. You go act on what it told you.
Claude Computer Use collapses that last step. The AI doesn’t just tell you what to do — it does it. That’s a different category of tool. It changes the calculus around trust, around delegation, around what kinds of tasks are worth a human’s time and attention.
We’re early. The safeguards aren’t perfect. The research preview label is there for a reason. But the direction is clear, and the pace of development in the first quarter of 2026 alone suggests Anthropic is not treating this as an experiment. They’re treating it as the product.
The era of AI that does rather than just answers has arrived. Whether you’re ready to hand over the keyboard is now the actual question.
