For the past three years, every major AI company has been racing to answer the same question: can we build an AI that does not just talk about work, but actually does it?
Anthropic just placed its biggest bet on that question. The company has launched a research preview that gives Claude the ability to directly control your Mac — clicking buttons, opening applications, typing into fields, scrolling through interfaces, and completing tasks on your behalf while you are away from your desk. Paired with a new mobile dispatch feature that turns your iPhone into a remote control for your computer, this is not an incremental update. It is a genuine attempt to shift Claude from a conversational tool into something closer to a digital operator.
This guide breaks down exactly how the feature works, what it can and cannot do right now, the real security risks you need to understand before enabling it, and what this release signals about where the AI agent race is heading in 2026.
What Just Launched — And Why It Is Different
Claude's computer use capability is available immediately as a research preview for paying subscribers — Claude Pro ($17/month) and Max ($100 or $200/month) — on macOS only. It is built into both Claude Cowork, Anthropic's agentic productivity tool, and Claude Code, its developer-focused command-line agent.
Alongside computer use, Anthropic also extended Dispatch — a feature that lets users assign Claude tasks from a mobile phone — into Claude Code for the first time. The combination creates an end-to-end pipeline that looks like this: you send an instruction from your iPhone, Claude executes it on your Mac desktop, and sends back the finished result. Your computer stays awake running the Claude app, your phone becomes the remote control, and you do not have to be anywhere near your desk for the work to happen.
That architecture — mobile instruction, desktop execution, asynchronous delivery — is what makes this release meaningfully different from anything Anthropic has shipped before. It is not about having a better chat experience. It is about Claude operating as a persistent background worker inside your actual computing environment.
How Computer Use Actually Works: The Priority System
When you assign Claude a task that requires computer interaction, it does not immediately reach for your mouse. Anthropic built a layered priority system that tries the most reliable path first and only escalates to screen-level control as a last resort.
Level 1 — Direct Connectors
Claude first checks whether a direct integration exists for the service involved. Connectors are available for Gmail, Google Drive, Slack, Google Calendar, and a growing list of other tools. When a connector is available, Claude uses it — pulling data through an API rather than navigating a user interface. This is fast, reliable, and far less error-prone than anything that follows.
Level 2 — Chrome Browser Navigation
If no connector exists but the task involves a web-based service, Claude navigates Chrome via Anthropic's Claude for Chrome browser extension. Still faster and more predictable than direct screen interaction, because the browser environment is more structured and consistent than a full desktop.
Level 3 — Direct Screen Control
Only when neither a connector nor a browser path is available does Claude fall back to actual screen interaction — taking screenshots of your desktop to understand what is visible, then clicking, typing, scrolling, and opening applications as a human operator would. Anthropic is candid in its own documentation: this mode is "the slowest and most error-prone" of the three. It is also the most powerful, because it can theoretically work with any application on your system.
This hierarchy is an honest acknowledgment of where the technology actually stands. Screen-level AI control is impressive and genuinely useful, but it is not yet reliable enough to be the first tool you reach for.
Dispatch: Your iPhone as a Remote for Your Mac
The feature that may ultimately have more practical impact than computer use itself is Dispatch — and it is worth understanding how it works.
Setup is straightforward: you pair your iPhone with your Mac by scanning a QR code inside the Claude app. From that point, you can send instructions to Claude from your phone at any time. Claude executes those instructions on your desktop — which must remain awake and running — and returns the results to your phone.
The use cases Anthropic envisions are a useful benchmark for what the system is designed for:
- Pulling your most important emails every morning into a structured briefing before you sit down at your desk
- Generating a weekly metrics report from local files and connected tools into a formatted document
- Organising a cluttered Downloads folder based on rules you define once
- Compiling competitive analysis from local and cloud-based sources into a deliverable you review rather than build
Scheduled tasks extend this further. Set a cadence — "every Friday afternoon," "every weekday at 7am" — and Claude handles the recurring work without you having to prompt it each time. One developer who tested the feature described the combination of Dispatch and scheduled tasks as effectively giving you "a background worker that can interact with any app on a cron job." That framing is accurate — and it captures why this is more significant than a feature update. It is the beginning of a model where AI operates in the background of your work rather than in response to your requests.
What Early Testing Actually Shows
Anthropic is calling this a research preview for a reason. The early hands-on results are mixed — and understanding where the feature works and where it does not is essential before building any real workflow dependency on it.
MacStories ran one of the first detailed evaluations. The results split cleanly. Claude successfully located specific screenshots on a Mac, summarised recent notes from a Notion database, listed notes saved on a given day, added a URL to Notion, summarised a recently received email, and recalled a screenshot from earlier in the session. These are all information retrieval and summarisation tasks — and they worked.
But it failed to open the Shortcuts app directly, send a screenshot via iMessage, list unfinished Todoist tasks due to an authorisation error, list active Terminal sessions, display a food order from an active Safari tab, or fetch a URL from Safari using AppleScript. The overall verdict from that evaluation: about a 50/50 success rate, too unreliable to depend on when you are away from your desk, but a genuine step forward.
The honest takeaway is that computer use works well for tasks that involve reading and summarising information, and struggles with tasks that require navigating complex multi-app workflows or interacting with applications that are not tightly integrated. That is consistent with where AI agents generally are right now — strong at the "understand and report" part, still rough at the "take complex action" part.
The Security Risks You Need to Understand
This is the section that requires careful reading before you enable any of these features.
Computer use runs outside the virtual machine that Cowork normally uses for file operations. That means Claude is interacting with your actual desktop and live applications — not an isolated sandbox. When Claude takes screenshots to understand your screen state, it can see anything visible on your display: personal documents, financial data, private messages, login credentials. Anthropic trains Claude to avoid engaging in stock trading, processing sensitive data, or capturing facial images — but the company's own documentation is explicit that "these guardrails are part of how Claude is trained and instructed, but they aren't absolute."
Anthropic has built meaningful layers of protection:
- Claude requests permission before accessing each application individually
- Sensitive categories of apps — investment platforms, cryptocurrency tools — are blocked by default
- Users can maintain a personal blocklist of applications Claude is never permitted to access
- The system actively scans for signs of prompt injection during computer use sessions
- Users can halt Claude at any point with a stop command
Despite these protections, Anthropic explicitly advises against using computer use for financial account management, legal documents, medical information, or any application containing other people's personal data. It also advises against using Cowork for workloads subject to HIPAA, FedRAMP, or FSI regulatory requirements.
There is an additional enterprise-specific concern that organisations need to take seriously. Cowork conversation history is stored locally on the user's device, not on Anthropic's servers — which is good for privacy. But enterprise features like audit logs, compliance APIs, and data exports do not currently capture Cowork activity. That means an organisation subject to regulatory oversight has no centralised record of what Claude did on a user's machine. For compliance-sensitive industries, that gap is not a minor limitation. It is potentially a dealbreaker for any formal deployment.
The deeper governance question — how do you distinguish between an action taken by a human and an action taken by an AI agent sharing the same mouse, keyboard, and screen — does not yet have a satisfying answer. As AI agents move from drafting text to taking real-world actions, that distinction becomes foundational for accountability and liability. It is an open problem, and it is worth being clear-eyed about that before granting Claude operator-level access to your system.
The Competitive Context: Why Anthropic Is Shipping Now
Anthropic's timing on this release is not coincidental. The company is entering a market that has been moving fast without it.
OpenClaw, the open-source framework for AI-driven computer control, went viral earlier in 2026 and created an entire ecosystem of derivative tools practically overnight. It proved that users wanted AI agents capable of taking real actions on their computers — and that they were willing to tolerate significant rough edges to get there. Nvidia entered the space with NemoClaw, its own framework aimed at making OpenClaw deployment simpler with added security controls. Smaller startups like Coasty are marketing polished desktop and browser automation tools directly to the same users Anthropic is targeting.
Meanwhile, Reuters has reported that OpenAI is actively courting enterprise customers in what it describes as a direct competitive battle with Anthropic — and that the ability to ship working agents has become the primary differentiator in that fight. The enterprise AI market is no longer being won on benchmark scores or model capability in the abstract. It is being won on which company can demonstrate that its AI actually operates reliably inside the tools organisations already use.
Anthropic is betting that its advantages — tighter consumer integration, an existing subscriber base, a more accessible setup experience, and a safety reputation that resonates with enterprise procurement — can compete against both the open-source community and well-funded rivals. Shipping now, even imperfectly, is how you establish the beachhead.
The Plugin Architecture: Where Enterprise Ambitions Become Clear
The longer-term strategic picture comes into focus when you look at how Anthropic is building out the plugin layer around computer use.
Plugins bundle skills, connectors, and sub-agents into a single install that turns Claude into a domain specialist for a specific function. Anthropic already lists plugins for legal workflows — contract review, NDA triage — finance — journal entries, reconciliation, variance analysis — and brand voice management. The vision is an AI agent that does not just have general computer access but is specifically trained, connected, and configured for the domain it operates in.
For enterprise buyers, this matters. A general-purpose computer-control agent is an impressive demo. A purpose-built legal assistant that reviews contracts, surfaces risks, and drafts amendments — operating inside the actual software your legal team already uses — is a product with a clear procurement case. The plugin architecture is how Anthropic converts the computer use capability into that kind of product.
Pricing and Availability
The feature is currently available to Claude Pro subscribers (from $17/month) and Max subscribers ($100 or $200/month). It is macOS only for now — Windows users are excluded from the research preview and have expressed predictable frustration.
One practical caveat worth noting: agentic tasks consume significantly more usage quota than standard chat interactions. Several Max subscribers have reported that Dispatch tasks burn through their monthly allowance at rates that surprised them. If you plan to use computer use or Dispatch heavily, factor that into your plan choice.
For teams, pricing starts at $20 per seat per month. Enterprise pricing is custom and includes admin controls to enable or disable Cowork capabilities across an organisation.
Should You Enable It Now?
The honest answer depends on your context.
If you are an individual knowledge worker comfortable with a 50 percent reliability rate on complex tasks and you understand what you are granting access to, this is worth experimenting with. Start with low-stakes information retrieval and summarisation tasks — the category where results are most consistent. Use the blocklist feature to explicitly exclude any application you would not want Claude to touch. Do not use it for anything involving financial accounts, sensitive personal data, or irreversible actions until the reliability improves significantly.
If you are evaluating this for enterprise deployment, the audit trail gap is the critical issue. Until Cowork activity is captured in the same compliance infrastructure as the rest of Microsoft 365 or your enterprise stack, formal deployment in a regulated environment carries real risk. Monitor Anthropic's documentation for when that gap closes.
For developers using Claude Code, the Dispatch integration opens genuinely interesting possibilities for background automation — but the known bug around multi-PDF payloads exceeding API limits is worth tracking before building any workflow that depends on document processing at scale.
The Bottom Line
Anthropic is not waiting for the technology to be perfect. It is shipping in public, with transparent acknowledgment of what does not work yet, betting that users will engage with a 50 percent success rate today in exchange for the trajectory it represents.
That is the correct bet at this stage of the market. The companies that establish user habits and workflow integrations now, even imperfectly, will have a structural advantage when reliability catches up to ambition. The failures today are recoverable — a missed click, a stalled task. The goal is to ensure they stay that way as the capabilities grow.
The AI industry spent three years proving machines can think. Anthropic is now testing a harder proposition: whether people are ready to let them act. Based on the early reception, the answer is a provisional yes — with permission dialogs, blocklists, and the reasonable expectation that nothing important gets deleted before the technology earns the trust it is asking for.
Have you tested Claude's computer use or Dispatch on your Mac yet? Drop a comment below with what you tried and how it performed — especially curious about any unexpected failures or surprising successes. Share this with someone who should know this capability exists.
No comments:
Post a Comment