Show HN: We Put Chromium on a Unikernel (OSS Apache 2.0)
8 days ago (github.com)
We’ve been building infrastructure to spin up browsers for AI agents. Originally, we built[0] it as a pool of warm Docker containers running Chromium, exposing:
- Chrome DevTools Protocol (for Playwright/Puppeteer)
- noVNC (for live view)
We’ve been following the unikernel space for a while, so we decided to see if we could get our image on one. We went with Unikraft Cloud[1]. Here’s how it did:
- Boot-up time: 10–20ms (vs. ~5s for Docker containers)
- Near 0 CPU/memory consumption when idle
- Still ~8GB RAM when active (headful Chromium)
Potential use cases:
- Standby mode during long-running jobs: unikernels can sleep after X sec of inactivity, reducing clock time costs
- Session reuse: auth/session cookies persist for hours/days. Basically as long as the cookies are valid
- Cold start speed: good for low-latency, event-based handling
We open sourced it with Apache 2.0! Feel free to fork or submit an issue / PR. Open to feedback or suggestions. www.github.com/onkernel/kernel-images
==
[0] https://github.com/onkernel/kernel-images
[2] Thanks to the Unikraft Cloud team @fhuici @nderjung @razvandeax for helping us figure this out (we're not affiliated)
[3] (OPs) @rgarcia @juecd
This looks excellent, and very fun to play with!
I used this one-liner to run an instance directly from remote source:
Be aware that the initial docker build is quite long... But caching kicks in for subsequent runs.
I look forward to playing with this!
I just wanna say the Dagger shell syntax is incredibly cool and I look forward to playing with it. I would not complain if you posted such self-promotional comments on any thread where it’s relevant because it goes to show the flexibility of the thing.
I saw your announcement post / many contentious HN comments, but wanted to chime in here with a supportive take, because I think time will reveal this design to be very compelling.
Thank you, I appreciate that!
I actually agonized over whether to include this particular one-liner, because of the risk of perceived self-promotion. In the end I decided to include it, because I actually used it, and I found it actually useful for anyone who wants to try Kernel with a one-liner. I made sure to not include a link to my project, to keep my karma in balance :)
I've been a fan of unikernels for a long time (we acquired the original Unikernel company at Docker), and I have to say applying it to browsers is genius. Now I'm surprised the unikernel community hasn't focused on this application sooner.
3 replies →
(OP) Happy to answer any questions! Some things we're still exploring:
- Mounting persistent storage for file i/o
- Replacing noVNC with a faster alternative
Would love feedback, especially if you’ve worked on fast-cold-start systems or unikernel-based infra.
How does fast-cold-start help here? What use cases did it enable that were previously not possible with a warm pool of Docker containers?
The big motivation we had for trying to do this on a unikernel is actually the session pause/resume. The fast cold restart is a nice addition (or put the other way, a super slow restart would prohibit session pause/resume from being useful)
This is probably a dumb question from someone who knows almost nothing about system engineering.
How hard would it be to boot a computer to this as an OS?
> How hard would it be to boot a computer to this as an OS?
Unikernels aren't meant to run as a bare metal OS on a standard computer like a PC. Instead they are applications wrapped in thin libraries that allow them to boot in hardware VM's provided by Intel vmx or AMD svm, etc. A hypervisor provides mechanisms for communication with hardware and other resources. They boot fast because the underlying system and hardware is already initialized and running.
The main idea of unikernels is to get rid of costly system calls like brk/sbrk called by malloc, open/read/write, etc. between the OS and application. The system never has to switch protection rings which saves a lot of time. This gives the application full control of its compute and memory resources with the possibility of direct hardware access depending on the host hardware and hypervisor. So you can attach things like NVM storage directly to the VM and let the application handle the disk and fs operations.
So to answer your original question of using such a wrapper to boot chrome on a PC: you will need a much, much bigger wrapper library which adds in all the hardware access which is a LOT of code (The GPU code alone is scary enough). You must also realize the fast boot time will be obliterated by hardware init which usually takes time as you have to jiggle certain hardware registers to wait, then probe again to see if things are working as advertised. This can take several seconds or more. In the end, you save nothing.
If you wanted an OS based on a hypervisor which boots unikernel applications you are at the mercy of the hardware to multiplex access or delegate that to a hypervisor adding more overhead. Again, you saved nothing.
In the end, your OS is really a CPU multiplexer and does a great job of providing all the primitives and resources in a generalized, uniform manner. I highly recommend reading this book: https://pages.cs.wisc.edu/~remzi/OSTEP/
> Unikernels aren't meant to run as a bare metal OS on a standard computer like a PC.
Originally this absolutely was one of the selling points. NetBSD's rumpkernel, for example, clearly was intended to support bare metal. In practice, though, unikernels are typically run within VMs, for the reason you hinted at--you usually still want a regular OS to multiplex your hardware and (unikernel-based) applications.
What about incus? Could incus provide some benefits as the host?
1 reply →
You would have to add support for the peripherals in the kernel, and have some kind of init system. You would also need a filesystem supported to boot the computer.
I was doing something similar for the entire OS a few years ago: cosmopolinux, a distribution of cosmopolitan binaries: https://github.com/csdvrx/cosmopolinux
My idea was to replace the WSL binaries to have a Linux distribution living on C:\, but that could also be booted baremetal if you didn't want to use Windows
I had to put together a multi stage init system for that: if you get the ISO, you can put in on a thumbdrive and boot it: https://gitlab.com/csdvrx/cosmopolinux
The only difference between them is the kernel and the filesystem: the github NTFS has a firecracker linux kernel, the gitlab ISO has a regular kernel with many modules.
I wanted to do a full NTFS solution but I couldn't find a bootloader I liked that would support booting from a NTFS partition.
Booting from an ISO was simpler and faster.
> I wanted to do a full NTFS solution but I couldn't find a bootloader I liked that would support booting from a NTFS partition.
Could you stick the Linux kernel and initramfs on the EFI boot partition as a UKI, and then just tell it about its rootfs being on the NTFS C drive? You don't really need any bootloader except the firmware's UEFI implementation on most modern PCs, and Linux supports NTFS.
3 replies →
When should I use this versus Browserbase or Browserless or Hyperbrowser?
Obviously since this is open source, then I can self host it. What other reasons?
Just curious!
Yeah, this is mostly just an OSS implementation for self-hosting! I'm not entirely sure what any of those companies use under the hood, in theory they could use this implementation if they wanted to gain the fast boot times, session reuse etc.
This is super cool. We’ve been looking into infra for AI agents. As others have noted, the difference in speed alone between docker and this is a huge win. Having our clients wait around for five seconds really adds up.
Awesome tech, excited to dig deeper for healthcare
Thank you!
Cool (and congrats on the demo)! Sounds like a promising approach. I work on browser use agents and one of the most difficult problems now is bot detection. Curious if you know how this impacts bot detection/fingerprinting?
Thank you! Yeah, the current implementation basically performs the same as a Docker container (i.e. not much). The interface is the same, so you can use BU/Playwright/Puppeteer's header configs to change as needed.
We did notice the unikernel cloud instances don't run into bot detection as often as our hosted docker instances, but I think that's mostly because Cloudflare haven't flagged Unikraft Cloud's IPs yet, hah.
Why detect "bots" when you can block them using captchas/proof-of-work solutions. Curious to know your use case for detecting bots.
I feel like their use case it's the other way around: avoid being detected as a bot.
1 reply →
This is awesome stuff. I think the biggest thing holding back a bunch agentic use cases is great infra and this is a a great step in the right direction. Love how fast it boots!
5s for Docker containers vs 20ms now ... holy moly, this is fast
From what we've seen, micro VMs could probably do something very fast too (150ms?) but we thought 20ms was pretty crazy.
Interesting work. It immediately brought boot2gecko to my mind. If I understand unikernels correctly - do you think it would be viable to run this on real HW?
I'm assuming the low latency cold starts are from a paused state, considering chrome itself takes a few seconds to boot? Or have you found some clever way to snapshot a running chrome and fork that?
Either way thanks for sharing.
It snapshots / pauses the entire unikernel instance after launching chromium, and then resumes the instance in <20ms with exactly the same state.
Is that safe? I was under the impression that snapshot/resume of ex. anything running crypto libraries was a minefield of duplicate keys and reused nonces.
Maybe you can clarify but this isn't actually using unikraft the kernel is it?
Maybe you can clarify that this is actually running a stripped down linux as unikraft does not have the support to run chrome itself.
> This unikernel implementation can only be run on Unikraft Cloud
Looking forward to one that'll run on my local machine, if I read this correctly?
Yeah, there are some nuances to Unikraft Cloud that their team did to get it running. I'll see if they can chime in and shed some light!
Wooow man this is crazy!! I wanna chat
What’s the size of the image?
Can this run inside a Lambda?
Haven't tried on a Lambda. Would be curious to know the results if you do!
It took us 8gb to get it up - maybe could be slimmed down if you took out the Anthropic Computer Use components, but browsers generally are pretty heavy.
Impressive work, @juecd!
Thank you!
why is the price bracket so weird? pro/10 gives each user less than half what individual plans would. It's almost like an attempt to dissuade pro purchases. Are there additional tooling you are not mentioning in the comparison?
2 replies →