Claude Can Now Use Your Computer. Here's the One Thing It's Still Missing.

TL;DR: Anthropic launched Claude computer use on March 23, 2026 — Claude can now autonomously control your Mac, open apps, click buttons, and complete entire workflows. It’s a genuine leap for AI agents. But there’s a critical missing piece: no way to record, share, or replay what Claude actually did. screencli is the answer — one command to capture a polished, shareable video of any Claude-driven browser session.

What Claude Computer Use Actually Is

On March 23, 2026, Anthropic shipped computer use into Claude Code and Claude Cowork for Pro and Max subscribers. This is not a minor feature update. Claude can now see your screen, move your cursor, click buttons, fill out forms, open files, and navigate any app on your Mac — all autonomously, even while you’re away from the machine.

The canonical demo: a user running late for a meeting texts Claude from their phone. “Export the pitch deck as a PDF and attach it to the meeting invite.” Claude opens the file, navigates the export dialog, finds the calendar event, and attaches the PDF. The user never touches the computer.

Anthropic’s own data suggests Claude Code users were already stretching the tool beyond coding — into general computer workflows. Computer use makes that official.

The technical picture:

• Claude captures screenshots of your screen at each step
• It interprets what it sees, decides the next action, and executes it
• It uses a permission-first approach — asks before accessing new apps
• Currently macOS only, in research preview
• Available to Claude Pro ($20/mo) and Claude Max ($100–$200/mo) subscribers
• Requires the beta header computer-use-2025-11-24 in API calls

This is meaningfully different from earlier computer-use experiments. It requires no setup, works across any Mac app, and it’s live for millions of subscribers right now.

The Gap Nobody Is Talking About

Here’s the problem: Claude does the work, and then… it’s gone.

You asked Claude to navigate your web app, demonstrate a feature, run through an onboarding flow, or test a form. Claude did it. But you have no recording. No shareable link. No replay. No proof of what happened — just a terminal log.

For a solo founder trying to ship a changelog video, a developer advocate demoing a new feature, or a QA engineer capturing a bug reproduction, this is a real gap. The agent did the work. You still need to capture and share it.

This isn’t a criticism of Anthropic — computer use is genuinely impressive for what it is. But the screen recording problem exists precisely because agents are now doing more things worth showing.

The three scenarios where this hurts most:

1. Product demos: Claude navigates your app to demonstrate a feature. You want a polished video for your changelog, launch post, or onboarding docs. You get nothing.
2. Bug reproductions: Claude reproduces a reported bug in the browser. Your QA team needs a recording to share with eng. You get nothing.
3. Agent showcases: You want to share what your Claude workflow actually does — to document it, post it, or impress a stakeholder. You get nothing.

The session happened. You just can’t show anyone.

How Claude Computer Use Compares to Browser-Based Agents

Most of the existing AI agent recording infrastructure is built around browser automation — Playwright-based agents that drive a browser window. Claude computer use is different: it operates at the OS level, controlling any app, not just the browser.

The implication: browser automation is still the right approach for headless, cross-platform, recorded agent sessions. Computer use is the right approach for ambient, unattended OS-level tasks. They’re not competing — they’re complementary.

For anything you want to record and share, browser-based agent sessions remain the better fit. Which brings us to the fix.

The Fix: screencli Records What Claude Drives

screencli is an open-source CLI that records AI-driven browser sessions as polished, shareable videos — with auto-zoom to action areas, click highlights, idle time trimming, and gradient backgrounds. One command. No screen recorder running in the background. No video editor.

The workflow that fills the Claude computer use gap:

# Instead of letting Claude computer use navigate your app silently,
# give Claude a browser session it can drive — and record it

npx screencli record --url https://your-app.com \
  --prompt "Navigate to the dashboard, filter by last 30 days, and export the CSV report" \
  --output demo.mp4

screencli launches a Playwright browser, sends the session to Claude’s agent loop, records every action, and produces a polished MP4. You get a composed video with:

• Auto-zoom to the area where Claude is clicking or typing
• Click highlights — animated rings on every interaction
• Cursor trails so viewers can follow Claude’s path
• Idle trim — dead air between actions is cut automatically
• Gradient background with rounded corners, no raw browser chrome

The output is a real demo video, not a raw screencast. Upload it, share the link, embed it in a changelog — done.

Using screencli from Claude Code directly

If you’re already using Claude Code, install the screencli skill and you can record demos from inside your Claude Code session:

npx skills add https://github.com/usefulagents/screencli --skill screencli

Then in Claude Code:

“Record a demo of the login flow on staging.myapp.com — show signing up, confirming email, and reaching the dashboard.”

Claude Code will use screencli to drive the browser, record the session, and return a shareable link. The entire thing runs in your terminal, no GUI required.

What This Looks Like in Practice

Three real workflows developers are using today:

1. Changelog demos Every time you ship a feature, run npx screencli record with a prompt describing the flow. screencli records a 30–60 second polished video. Paste the link in your changelog post. No retakes, no editing, no screen recorder app.

2. Bug reproductions Describe the reproduction steps as a prompt. screencli drives the browser through the exact steps, records it, and gives you a shareable link to attach to the GitHub issue. QA and eng see the same thing — a clean replay.

3. Onboarding walkthroughs Record a polished walkthrough of your app’s key flows once. When the UI changes, re-run the same command. Always-fresh documentation videos with zero manual effort.

The Broader Pattern

Claude computer use is a preview of where agent tooling is going: less setup, more ambient, deeper OS integration. The agents will do more and more. The gap between “the agent did it” and “here’s a shareable record of what the agent did” is only going to grow.

This is exactly why screencli exists. Agents don’t just need to do things — they need to show things. The recording layer is infrastructure, not an afterthought.

The current state: Claude computer use is live, impressive, and useful. For browser-based workflows you want to capture and share, screencli remains the most direct path from “I described what I wanted” to “here’s the polished video.”

FAQ

Does screencli work with Claude computer use directly? Not directly — Claude computer use operates at the macOS level and doesn’t expose a recording hook. screencli works with Playwright-based browser sessions, which is the right choice for anything you want to record. If your workflow involves a browser, screencli is the better fit.

Can I use screencli without Claude Code? Yes. screencli works standalone: npx screencli record --url <url> --prompt "<prompt>". If you have an ANTHROPIC_API_KEY set, it calls Claude directly. If you’re logged into the cloud, it proxies through screencli’s servers.

What’s the difference between screencli and just using Playwright’s built-in video recording? Playwright records raw video. screencli records + post-processes: auto-zoom to action areas, click highlights, cursor trails, idle time trimming, gradient background with padding and rounded corners. The output is a demo-ready MP4, not a raw capture.

Does screencli support recording in CI/CD pipelines? Yes — screencli runs headless and can be triggered from any CI environment. It’s designed to be scriptable and automation-friendly.

How much does screencli cost? screencli is open-source and free to run locally. The cloud plan (for hosted video storage and sharing links) starts at $12/month. The CLI itself is MIT-licensed and free forever.