µSandbox

2/ after a week of benchmarking and revisiting the fs implementation, our new benchmarks outperform docker's in quite a number of cases we measured🎉

“I haven’t needed a sandbox so far.” That’s the most common response we hear. But as soon as you let AI agents execute code, call tools, or access real systems, you’ve already changed the risk model.

there’s so so much on the way. can’t really tell y’all how excited i am. next week is gonna be even much better. 😊🤗 sandboxes can be so much better, and we really want to make it as seamless as possible. anyways, good night fellow nerds. need to go take a rest ahead of next week. there’s a lot we’re cooking, and you need to get served properly … you just gonna need be at the dinner table. 🧑🏾‍🍳

⚡️microsandbox filesystem is now 20-62% faster than docker’s! ⚡️ this is from being 100x slower a day ago.

4/ if you didn't know, microsandbox is the easiest way to sandbox AI agents locally. blog post on how we pulled this off coming soon. stay tuned 📻 there's more on the way. try it out today at: microsandbox.dev

3/ there's still more to squeeze out, and we're just getting started. expect these improvements to land in v0.4.0 (late next week). benchmark code: github.com/superradcompan…

a late week win 🔥🧵 1/ a community member on discord flagged that our file system implementation was 10x slower than docker's. in some areas, even 100x slower. we expected this with the recent rewrite of microsandbox. optimizing it was always on short term roadmap. well, with a few ideas off our sleeves, we got to work.

Microsandbox (YC F26) Boots a real VM in under 100ms, no daemon, no server, no root. libkrun microVMs with hardware-level isolation. SDK is three lines of code. Secrets never enter the guest. Apache 2.0. microsandbox.dev Founders: @theappcypher

the AI landscape has and will continue changing. but i think one thing that could increase ones chances of winning would be flexibility and speed to adapt to change. one example is the sharp realisation that we can't inject secrets into sandboxes due to prompt injection as a means to exfiltrate keys. same applies to sudden realisation with file system, etc. big ships built without extensibility will be hard to sail the other direction; while the ones baked in with extensibility will adapt, and immediately thrive once they have that network effect. the latter is the stance we're taking @microsandbox. don't try to be smart by predicting the future. be wise in being well positioned. the best we can do is do the hard-work upfront for our ships to be extensible/adaptable. it's the reason most of the sudden realisations lately haven't taken i and @theappcypher by surprise. make every component of the sandbox plug-and-play native ... and i mean, everything! i'm glad we're almost out of the weeds on that initial ground work. better putting us in a position to adapt and steer the ship as the industry realise what works and doesn't. we do have a lot of work to do on the network effect, but we anticipate this will grow as we become known for steering sharply into the right direction and sailing quickly.

@freddie_spirit The best for flawless and secure agentic Ai

@microsandbox seems amazing for building agent sandboxes! good job!

this is the core premise around how we rebuilt @microsandbox. expose the file system so devs can write plug-and-play filesystem that abstracts some services. map IAM policies to file permissions, etc. we’ve got the first leg of this landed in core. it should land in our SDK soon ;) it’s the same core principle of foreign data types in postgres.

some of those already landed, but there's so so much more to come. if you're running agentic workloads at scale, feel free to shoot us a message. we'd love to help.

Coming weeks are gonna be so much fun! Can't wait to share all the cool things we've been working on in the background, and how we're rethinking sandboxes for the agentic era. Insider tip: see website + github project, or discord 🤭

microsandbox: self-hosted platform for running untrusted code with isolation

🚀 Game-changer for AI developers! @microsandbox is the Code Interpreting platform we've been waiting for. 🤖 Perfect for AI Code Execution: • Safe execution of AI-generated code • Code interpreting for AI applications • LLM code output sandboxing • AI agent script testing ✨ Why it's revolutionary: • Hardware-level VM isolation with microVMs • Lightning-fast startup (<200ms) • Self-hosted alternative to E2B • Complete control over your AI infrastructure 🔒 No more worrying about: ❌ Malicious AI-generated code ❌ System compromise from experimental scripts ❌ Vendor lock-in with cloud sandboxes This is exactly what the AI community needs for safe code interpreting! 🔗 github.com/microsandbox/m… #AI #LLM #CodeInterpreting #E2B #OpenSource #AIAgents #Security

Reached 1k stars! 🎉 That’s it. Time to get back to work.

Look ma, I made it to HN frontpage! #44136867" target="_blank" rel="nofollow noopener">news.ycombinator.com/item?id=441359… #hackernews #microsandbox

Introducing Microsandbox [📦] A self-hosted solution that bootstraps fast lightweight VMs where you can run AI-generated or untrusted user code. You can try it today on macOS or Linux.