Mantis Ring
A UI-fixing agent whose fixes are re-measured, not trusted. Sandboxed on Vultr.
- Vultr VX1 · Seattle
- Vultr Serverless Inference
- gVisor sandbox
Judges: test in 60 seconds
-
See a finished run The latest verified run: its receipt, before-and-after screenshots, the models and their cost, and the patch.Latest verified run
-
Run the sample yourself Live runs open soon.
-
Run it on a public GitHub repository One app at the root of the repository, up to 300 MB, that installs from npm and needs no keys or database.
-
Check what is running Sandbox runtime, model service, self-test results and today's spend, checked live or read from the last real self-test.Status
How a run works
- Clone
- Install (proxy)
- Look
- Fix
- Gate
- Vision check
- Approve
Ring is a coding agent that fixes UI bugs in a web project: text that overflows on a phone, layouts that break in another language, colours that disappear in dark mode. It works on a copy of the project inside a throwaway gVisor sandbox on Vultr, with no secrets and no way out except a proxy that allows only npm and GitHub. After each batch of edits, Mantis, an open-source checker that uses no AI, measures every route, screen size, colour scheme and language in the task again and compares them with the starting point. A change that only hides a problem is reported as a cover-up and rolled back, and so is one that makes a page stop loading. Each run ends with a receipt that shows what was fixed, what is new, which models worked on it and what they cost. A person approves before the patch is handed out as a file; the demo never applies it anywhere.