May 6, 2026
Ctrl+Z for AI chaos?
Show HN: Tilde.run – Agent Sandbox with a Transactional, Versioned Filesystem
AI’s ‘undo button’ lands, and the comments instantly turn into a trust battle
TLDR: Tilde.run wants to make AI assistants safer by giving their actions an “undo” history across your files and apps. Commenters liked that idea, but the real debate was whether it truly protects you in the messy real world — and whether anyone trusts yet another shiny AI tool.
A new Show HN project called Tilde.run is pitching a very comforting idea in a very chaotic AI moment: let an AI helper do work with your files and services, but keep everything inside a trackable sandbox so you can rewind mistakes. In plain English, it’s selling the dream of an AI that can touch real work stuff without turning your data into a horror story. And yes, some people were immediately into it. One early reaction said the versioned storage sandbox is the real magic here — basically, the feature that makes this feel different from the endless stream of “AI agent” tools flooding the internet.
But the comments did what comments do best: they threw cold water everywhere. One skeptic cut straight to the heart of the pitch: if the AI only changes files, sure, you can roll that back — but if it sends emails, edits outside systems, or does anything in the real world, then an “undo” button may not save you. That sparked the big tension of the thread: is this real safety, or just neat-looking safety?
Then came the trust drama. Another commenter side-eyed the hosted version, pointing out that the homepage asks for analytics while only the command-line tool and developer kit are open source. Translation for non-geeks: people like the idea, but they’re not eager to hand over the keys. And in the funniest grumpy subplot, one user basically said they’ve developed “AI landing page fatigue” — too many glossy demos, too many animations, too many promises. In other words: cool concept, but the crowd wants receipts, not vibes.
Key Points
- •Tilde.run is presented as an agent sandbox for autonomous code execution.
- •Each agent run is treated as a transaction that can be rolled back.
- •The system combines code from GitHub, data from S3, and documents from Drive into a single versioned filesystem.
- •Every outbound call is checked and logged.
- •The product is positioned as a way to use autonomous code more safely with real data.