Launching today
CreateOS Sandbox
Instant, hardware Isolated Sandboxes for AI agents
207 followers
Instant, hardware Isolated Sandboxes for AI agents
207 followers
CreateOS Sandbox gives AI agent builders their own fast, secure, hardware isolated, sandbox in ~30ms (p90). We have suit of CLI, SDK, 50+ SDK real world examples, claude plugins, computeSDK integration, and more
The 30ms cold start is genuinely impressive for hardware isolated sandboxes. Whoever tuned that cold boot path clearly obsessed over it.
CreateOS
@tunahancangelir Correctly clocked. Prewarmed snapshots + a dedicated boot lane so new-VM spawns don't queue behind everything else — took a while to get the tail latency down, not just the median. Good to hear it holds up outside our own benchmarks.
CreateOS
Hey everyone, Mrudul here from CreateOS. I spend most of my time talking to companies that want to run AI‑generated, untrusted code in production, and the same obstacle keeps appearing. It’s rarely about whether the code runs; it’s about what the code can reach, where the data resides, and who can see it.
That’s the problem I care about with CreateOS Sandbox. Each sandbox is an isolated environment, and the most useful feature is the ability to create a private network across many of them. You can stand up a cluster of agents that communicate over private DNS while remaining sealed from everything else, something most teams assume requires a full VPC and weeks of setup.
The other key advantage is control. You can run the entire stack on your own hardware, including the control plane and storage, and we are SOC 2 Type II and ISO 27001 certified. For teams undergoing a serious security review, that distinction can be the difference between a trial and a polite “no.”
CreateOS
@mrudul_gole This is the conversation on repeat, exactly. Most sandbox tools solve "does the code run" and stop, the private-network-per-cluster piece was the part we debated longest internally, because it's easy to fake with DNS tricks and hard to do right with actual per-node isolation. On the compliance side, self-hosted control plane + SOC 2/ISO wasn't a checkbox exercise -> it's literally what gets us past the security review instead of a polite no. Happy to go deeper on the private-networking design if anyone wants internals.
CreateOS Sandbox
Super proud of the entire team for shipping this 🤩🎉
As the marketing lead here at CreateOS, seeing all the hard work go into building CreateOS Sandbox for secure, high-performance AI agent isolation has been incredible.
Can't wait to see what amazing agents developers build with sub-30ms provisioning and bulletproof eBPF egress control. Let us know what you think.
CreateOS
Team effort top to bottom -> kernel work, gateway, scheduler, docs, everything. Sub-30ms provisioning was the constraint we refused to break, and it shaped nearly every decision below it. Appreciate you rallying launch day. Now let's see what people build with it.
CreateOS
If your AI agents run code you don't fully trust, this is worth a look. Each agent gets its own isolated sandbox with a real guest kernel, ready in about 30ms, so isolation stops being a tradeoff against speed.
The smarter part is the egress control. Rules are enforced in kernel via eBPF, from outside the sandbox, so even a fully compromised agent can't get around your allowlist. Add in millisecond sandbox forking, your own S3 as shared storage, and the option to run it on your own hardware, and it covers most of what actually breaks in production agent setups.
There's already a Claude plugin so Claude can run untrusted code in a sandbox that wipes itself when idle, plus 50+ real examples to build from. 500 free alpha credits, no card needed, so it costs nothing to test against your own workload.
CreateOS
@navedux There is a sandbox feature in Claude-Code and Codex, but apparently it's a process level and has a lot of gotchas, and you won't feel comfortable. That's where you can utilize sandboxes with your existing cloud or Codex harness to run any kind of untrusted nodes and do all kinds of testing, benchmarking, and development.
So if you want to take it to one more level up, you can utilize SDK and create your own agent. Also you can use Anthropic-managed agents with self-hosted sandboxes.
Hardware-isolated sandboxes are a strong direction for agent work. The practical question I always look for is whether the environment makes rollback, artifacts, network access, and final-state evidence obvious enough that a small team can trust the result without babysitting it.
CreateOS
@krekeltronics That I consider application layer not the infra layer but you can use fork smartly and figure it out
Spun up a sandbox in about 30 seconds and the hardware isolation gave me real peace of mind for running untrusted agent code. The CLI felt snappy and the SDK examples made it easy to wire into my existing flow without much fuss.
CreateOS
@tugayabaylungn This is exactly the reaction we built for — untrusted code, real kernel boundary, no crossed fingers. Glad the SDK examples got you moving without a fight. Curious what you end up wiring it into.
Spun up a sandbox in around 30ms like they claim, the CLI felt snappy and the SDK examples actually made sense for once.
CreateOS
@emircansowj Appreciate the backhanded compliment to the rest of the industry's docs. Glad it actually clicked instead of fighting you.