HyperFrames
Write a video as HTML, CSS and animation, render it to MP4 with a headless browser. Built so a coding agent can do the writing.
Install it
Safe?Works?Worth it?
Every week a new AI tool goes viral. We read its code, run it in a locked-down sandbox, give it real work, and tell you whether to install it.
- $ sv quarantine trending/agent-tool
- static review0 critical, 2 warnings
- networkoff
- installok in 14 s
- task 1: refactor a modulepassed
- task 2: fix a real bugfailed
- cost$0.41
- Safe
- 4/5
- Works
- 3/5
- Worth it
- 3/5
Latest verdicts
Every verdict comes with the test log and the video, plus what's still in the sandbox.
Scrapling
A Python scraping framework with fetchers, a browser, spiders, an MCP server and elements that try to keep working after a site's layout changes.
Use with care
Cloudflare security-audit-skill
An agent skill that splits a security audit into six phases, and makes a different agent try to disprove every bug before it is reported.
Verdict pendingVideo Dec 1
How a tool gets a verdict
The striped line is the sandbox wall. Untrusted code only ever runs inside it.
- 1
Read the code first
Install scripts, dependencies, hidden network calls, look-alike packages. If something looks wrong, it never runs. It goes to quarantine.
- Inside the sandbox2
Install it, locked down
Docker or a throwaway server. Network off wherever possible, no real keys. Never on our own machine.
3Give it real work
Tasks a real user would give it, recorded on screen. We count what passed, what failed and what it cost.
- 4
Write the verdict
Safe, Works and Worth it, each out of 5. Then one call: install it, use it with care, or skip it.
Rules we don't break
- Unknown publisher, no execution. Code from people we can't verify gets a static review only. If it fails, it lands in quarantine with the reason.
- Nothing runs on a real machine. Every install happens in a sandbox we can throw away.
- Nobody pays for a verdict. Affiliate links are labelled and never change a score.
- You see the real output. If a test fails, you watch it fail.
Two verdicts a week
Tuesday: one tool, tested in full. Friday: the five most viral tools of the week, checked fast. Want a repo checked? Suggest it and vote.