toolbox
intermediate
Claude
Claude Computer Use: Let AI Run Your Desktop for You
A practical guide to Claude Computer Use: what it automates, how to set it up safely, and the patterns that actually work for professionals in Asia.
9 min read24 April 2026
ai-agents
automation
productivity
claude
desktop
asia
mcp
computer-use
Claude Computer Use lets Anthropic's models see your screen and click, type, and scroll on your behalf, scoring 72.5% on the OSWorld desktop benchmark.
You reach it through the Claude desktop app on macOS and Windows (Pro from about US$20 per month, Max at US$200) or through the API for developers building custom agents.
It shines on legacy apps, cross-system data entry, and outcome-style prompts; it struggles with speed-critical work and needs a careful permission setup to stay safe.
Why This Matters
For years the promise of AI agents was mostly a demo video. That changed in 2026. Claude Computer Use, first shown in preview in late 2024, jumped from under 15% on the OSWorld desktop benchmark to 72.5% by March 2026, according to Anthropic's own reporting covered in industry briefings. In practical terms, that means the model is now useful enough to handle real chores: drafting a weekly expense sheet, triaging support tickets, or pulling Salesforce data into a Notion doc while you are in a meeting.
The reason this matters for Asian professionals is that much of the region's software stack is stubbornly old. Finance teams in Jakarta still wrestle with desktop CRMs. HR systems in Manila often lack modern APIs. A tool that can just look at a screen and click buttons like a person can bridge a lot of that gap, especially for small and medium businesses that cannot afford a full integration team. Add in local constraints like patchy internet and staff working across mobile and desktop, and the appeal of an AI that can run tasks overnight on your machine becomes obvious.
The catch: this is not magic. It is slow, it makes mistakes, and it can do real damage if you grant too much access. The rest of this guide walks through how to set it up properly, what to automate first, and the patterns that separate a useful assistant from a liability.
The reason this matters for Asian professionals is that much of the region's software stack is stubbornly old. Finance teams in Jakarta still wrestle with desktop CRMs. HR systems in Manila often lack modern APIs. A tool that can just look at a screen and click buttons like a person can bridge a lot of that gap, especially for small and medium businesses that cannot afford a full integration team. Add in local constraints like patchy internet and staff working across mobile and desktop, and the appeal of an AI that can run tasks overnight on your machine becomes obvious.
The catch: this is not magic. It is slow, it makes mistakes, and it can do real damage if you grant too much access. The rest of this guide walks through how to set it up properly, what to automate first, and the patterns that separate a useful assistant from a liability.
How to Do It
1
Check you are on the right Claude plan
Claude Computer Use is available on the Pro and Max consumer plans through the Claude desktop app, and through the API for developers. Pro sits at roughly US$20 per month; Max at around US$200 per month with higher limits and role-based controls for teams. Download the macOS or Windows app from Claude.ai. Linux is not officially supported at the time of writing. If you are already on the Free tier, you can test most of what is covered here by upgrading for a single month before committing.
2
Turn on Cowork mode and enable desktop control
Open the desktop app, head to Settings, and enable Cowork mode. This is the surface where Claude can see your screen, read files, and execute actions. Computer Use itself is off by default: you must toggle it on and approve the apps you want Claude to touch. Start by granting access to one low-stakes application, such as a web browser or a scratch folder. Do not give it your banking app or your production terminal on day one.
3
Install a handful of MCP connectors before using visual control
The Model Context Protocol is Anthropic's open standard for connecting Claude to structured tools. There are MCP servers for Slack, GitHub, Notion, Google Drive, and thousands more. Always prefer an MCP over screen clicking: it is faster, more reliable, and less error prone. From Settings, add the MCPs that match your daily tools, then authenticate each one with the account you actually want Claude to act as.
4
Write a CLAUDE.md style context file
Before you give Claude a real task, spend ten minutes writing a short context document. Include your role, your company, your tone of voice, any terminology it should use, and any rules (for example, never send external emails without asking). Drop this into a Project or reference it in your prompt. Skipping this step is the single biggest source of wasted runs: every task ends up starting with twenty minutes of re-explaining who you are.
5
Start with a low-stakes test task
Pick something boring and reversible for your first run. A good example: 'Open my Downloads folder, find all PDFs older than 30 days, move them into a new folder called Archive, and write a short summary of what you moved.' Watch Claude plan the task, approve each step, and see how it recovers from small errors. Resist the temptation to walk away until you have done this three or four times and trust the behaviour.
6
Shift to outcome-style prompts, not step-by-step scripts
Claude Computer Use performs noticeably better when you describe the outcome you want rather than each click. Compare 'Click File, then Save As, then type report.pdf' with 'Export this document as a PDF called report.pdf and save it to my Desktop.' The second is shorter, more tolerant of UI changes, and lets Claude pick the best path. Use step-by-step language only for edge cases where the happy path is ambiguous.
7
Use Dispatch to hand off between phone and desktop
Dispatch is the feature that lets you queue a task from your phone and have your desktop run it. This is useful for the Asia commute: on the MRT in Singapore or the KTM in Kuala Lumpur, you can send Claude a brief and the work is waiting when you sit down. Pair your devices once via the QR code flow, then schedule tasks to run during off-hours if you want to save on token costs on the Pro plan.
8
Review, then tighten permissions
After a week of use, audit the apps Claude can access. Remove anything it has not actually needed. For anything sensitive, consider leaving Remote Control on but Computer Use off, so Claude can observe but not take destructive actions without an extra confirmation. If you use Claude Code for engineering work, enable its sandboxing features (PID-namespace isolation and script caps) before letting it touch production.
What This Actually Looks Like
The Prompt
You are my operations assistant. Every Monday at 9am local time, build a weekly revenue brief: pull this week's sales totals from the Salesforce web UI (login is saved in my browser), pull last week's totals for comparison, add a Slack #cs-tickets summary of open P1 issues, and save everything to a Google Doc titled 'Revenue Brief - [week number]'. Share the doc in the #leadership Slack channel with a one-paragraph summary. Ask me before anything is shared externally.
Example output — your results will vary based on your inputs
Claude generates a five-step plan: (1) open Chrome and navigate to the Salesforce dashboard; (2) capture current-week totals by region; (3) fetch last week's totals from the same report; (4) use the Slack MCP to count P1 tickets from #cs-tickets in the last seven days; (5) create a new Google Doc, paste a formatted brief, and post to #leadership. It pauses at step 5 for approval before sharing externally. Total runtime: around 6 to 8 minutes. It flags that two Salesforce rows had blank region codes and asks whether to exclude or estimate them.
How to Edit This
The first run will almost always surface one or two assumptions Claude made that you would prefer to correct: which week-number convention to use (ISO or local), whether to include cancelled deals, or how to caveat missing data. Treat the first run as a calibration exercise; once the prompt is refined, you can schedule it weekly and spend under 60 seconds reviewing the output.
Common Mistakes
Granting access to everything at once
Micromanaging the clicks
Skipping the context file
Treating it like a fast script
Not pairing Dispatch
Tools That Work for This
Claude desktop app— The primary entry point for Computer Use on macOS and Windows. Pro (US$20/month) is enough to evaluate; Max (US$200/month) makes sense once you depend on it daily.
Claude Code— Anthropic's terminal-based coding agent, which pairs naturally with Computer Use for engineering workflows and can run background tasks inside sandboxes.
Model Context Protocol (MCP)— The open protocol behind Claude's connector library. Browse community-built MCP servers for Slack, Notion, GitHub, and thousands of other tools.
Notion— A reliable target for generated briefs, summaries, and research notes; the Notion MCP is one of the more mature integrations.
Slack— If your team already lives in Slack, the Slack MCP removes most reasons to use visual control for messaging workflows.
1Password— A sensible way to manage the credentials you do not want Claude to see. Use browser autofill rather than pasting secrets into prompts.
What Claude Computer Use actually does
At its core, Claude Computer Use is vision plus action. The model looks at a screenshot of your desktop, decides what to click or type, performs the action, and then looks again. It repeats that loop until the outcome you asked for is reached. In Anthropic's own OSWorld benchmark, which tests real desktop tasks across file managers, browsers, and productivity apps, Claude rose from under 15% accuracy in late 2024 to 72.5% by March 2026. That is the threshold where the tool moves from impressive demo to something you can actually trust with your own work.
Where it fits for Asian professionals
For teams across Southeast Asia, the practical value is highest in three places. The first is legacy software: the desktop accounting package that has no API, the local CRM that never got a proper integration, the government portal you log into once a month. The second is cross-system grunt work: copying numbers from Salesforce to a Google Doc, triaging Slack tickets into Jira, archiving PDFs from email into a Drive folder. The third is overnight or unattended tasks, where the model can work while you sleep or commute. If your day is mostly in a browser, Chrome-based agents may suit you better; if it is spread across native desktop apps, Claude has the clearest edge today.
What to watch before you hand it the keys
Two cautions. First, speed: Computer Use is deliberate by design. Tasks that would take a well-written script a few seconds may take Claude several minutes. This is fine for batch work and a bad fit for anything time-critical. Second, permissions: by default Claude cannot touch any app you have not approved. Keep it that way. Grant access app by app, keep destructive actions behind a manual confirmation, and never paste banking credentials into a prompt. Used with care, it is a productivity lift. Used carelessly, it is a liability.
Frequently Asked Questions
With sensible permissions, yes for low-to-medium-risk tasks. Grant access app by app, keep destructive actions behind a confirmation step, and never give it credentials for payment systems. For anything truly sensitive, keep it on observe-only mode while you build trust.
Yes. Claude handles Bahasa Indonesia, Malay, Thai, Vietnamese, Tagalog, Mandarin, Japanese, Korean, and Hindi well. In practice, many Asian professionals prompt in English and ask for local-language outputs; this tends to produce the most reliable planning.
Operator and Mariner lean more towards in-browser workflows; Claude Computer Use covers the full desktop, including native apps and legacy software. If your work is mostly in a browser, all three are worth trying. If you need to automate a desktop accounting package, Claude is currently the clearest choice.
No. For anything where a real API exists, Zapier, n8n, or a proper script will always be faster, cheaper, and more reliable. Computer Use earns its keep where APIs are missing, brittle, or behind a login you cannot automate.
On the Pro plan the cost is rolled into your subscription, subject to usage limits. On the API, pricing follows the underlying model: Sonnet 4 is roughly US$3 per million input tokens and US$15 per million output at the time of writing. A typical 5-minute automated task consumes a few cents of tokens.
Next Steps
Once you are comfortable with a handful of Cowork tasks, read our guide on Context Engineering to get more out of every prompt, and the AI Agent Prompts guide for ideas you can port to Claude. If you work in Southeast Asia, our AI Regulation Across Asia overview is worth a look before you let an agent touch customer data.