I Let AI Take Over My Computer... Here's What Happened
TL;DR
Anthropic’s new combo is basically “text your computer and it does the work” — Oliver frames Claude’s computer use plus Dispatch as a remote-control layer for real desktop tasks like pulling reports, summarizing files, posting to Slack, and acting from your phone while you’re away from your desk.
Computer use means Claude can literally operate your desktop like a person — it can see the screen, move the cursor, click through apps, and work without a special API, which Oliver summarizes as: if a human can do it on a desktop, Claude can too.
Dispatch is the orchestration piece that decides how to complete the task — from the phone, it can route work through Claude code, Claude co-work, computer use, or existing integrations with apps like Slack, Google Calendar, and Chrome.
The setup is gated but flexible: Apple-only for now, Claude Pro or Max, with lots of permissions and confirmations — users need the Claude desktop app plus phone app, then must enable screen recording, accessibility, file access, and choose app/folder boundaries so Claude only acts inside defined limits.
In a real test, Claude analyzed YouTube thumbnail and title A/B tests directly from screenshots on Oliver’s machine — it concluded that “how” framing, specific details like dollar amounts or guest names, and strong branding such as “This Week in AI” or episode numbers were common traits among winning variants, while positive framing beat more dramatic, doomer-style titles.
Oliver’s bigger takeaway is Anthropic’s product speed, not just the demo — he argues the company has been shipping aggressively since OpenClaw launched, citing Claude Co-work, Opus 4.6, Sonnet 4.6, and these desktop-control features as evidence Anthropic has “leapfrogged OpenAI” in many people’s minds.
The Breakdown
The pitch: text your computer like it’s an assistant
Oliver opens with the dream scenario: message your computer from anywhere — lunch, the beach, wherever — and have it actually do the task, not just remind you about it. He positions Anthropic’s new computer use and Dispatch as exactly that, while joking that Anthropic is basically copying OpenClaw and productizing it.
Computer use and Dispatch, separated cleanly
He explains computer use as Claude literally operating your machine: seeing the screen, moving the cursor, clicking buttons, and navigating apps “exactly how a person would.” Then he contrasts that with Dispatch, which is the bridge from your phone to your computer and the decision-maker that figures out whether to use Claude co-work, Claude code, computer use, or an existing integration like Slack, Google Calendar, or Chrome.
Setup is straightforward, but the permissions are serious
To get started, you install the Claude desktop app, connect your account, and — for now — this only works on Apple devices with Claude Pro or Max. Then comes the important part: enabling computer use, screen recording, accessibility, file access, and defining exactly which apps and folders Claude can touch, with constant confirmation prompts that Oliver admits are a little annoying but clearly better than letting an agent roam freely.
First real test: Claude digs through screenshot files to analyze YouTube A/B tests
Oliver tries this on his own workflow: he asks Claude from his phone to look through recent screenshots of thumbnail tests and explain why some performed better. The process feels real, not magical — Claude starts searching through Desktop and Downloads, asks for permission, takes too long, grabs files he didn’t want, and only gets on track after he clarifies, “just use the 4 AB test.”
The actual analysis was useful, and pretty specific
Once focused, Claude returns a breakdown Oliver seems genuinely impressed by. It notices that three of the four tests were really title A/B tests, then pulls out patterns: titles starting with “how” consistently won, specificity made promises feel more credible, and branding like “This Week in AI” or an episode number appeared in every winner; most surprisingly, positive framing beat more dramatic, doomer-style language.
Second demo: posting to X from Chrome, controlled remotely
Next he tries something simpler and more theatrical: telling Dispatch to open X in Chrome and post, “Hey, this is Oliver’s computer posting this using computer use from Claude.” Claude moves quickly, sends back a screenshot showing the draft ready to go, and after Oliver confirms from his phone, it posts successfully — a nice proof point for road-warrior or locked-out-account use cases.
The bigger story is Anthropic’s release pace
He closes less on the feature itself and more on what it signals about Anthropic. Since late January or early February, after OpenClaw appeared, he says the company has been on a tear with Claude Co-work, Opus 4.6, Sonnet 4.6, and a stream of new features, arguing that in many people’s minds Anthropic has now leapfrogged OpenAI as the top model provider.