Inference and eval
Let an agent serve a local model, run an eval harness, or sweep quantization settings — in a box, not on your workstation while you're trying to use it.
Most managed agent sandboxes are CPU-only. Attach one GPU — or several — to an isolated
Linux box on hardware you already own, and the agent inside gets
nvidia-smi,
CUDA, and the whole device.
# pre-flight: is passthrough actually working on this host?
$ containarium backends validate-gpu
ok · NVIDIA · driver 550.x · nvidia-smi ran inside a throwaway box
# one GPU, by device index
$ containarium create infer-box --gpu 0
# or several, for multi-GPU work — index or PCI address
$ containarium create train-box --gpu 0 --gpu 1
$ containarium create train-box --gpu 0000:0b:00.0The gap
Plenty of agent work is CPU-shaped: edit files, run a build, run tests. But the moment the task is "benchmark this kernel", "fine-tune on our data", or "reproduce the training bug", a CPU-only sandbox is the wrong tool and the agent stalls.
Let an agent serve a local model, run an eval harness, or sweep quantization settings — in a box, not on your workstation while you're trying to use it.
Kernel and driver-adjacent code can't be tested by inspection. The agent needs a device present to know whether its change builds and runs.
A box persists, so a long run survives the agent disconnecting — and the dataset and checkpoints stay on your disk rather than being re-uploaded each time.
Your hardware
If there is a GPU in the office, in a rack, or under your desk, it can be agent capacity this afternoon. Self-hosting turns a card you already bought into sandbox capacity at close to zero marginal cost — and nothing about the workload has to leave your network.
A workstation card sitting idle overnight is the cheapest capacity you will ever get. The comparison that matters isn't rate-vs-rate, it's rented-vs-already-owned.
Proprietary weights, customer data, and anything under a data-residency rule never leave the machine. That is often the reason a hosted sandbox was never an option.
The core is open source and installs on a single Ubuntu host. No vendor in the path between your agent and your GPU.
Each box is its own environment, so one agent's CUDA toolkit or driver-version pin can't break another's. We run five boxes ourselves.
How it works
Install on an Ubuntu machine with an NVIDIA card and driver. containarium backends validate-gpu launches a throwaway container, runs nvidia-smi inside it, tears it down, and tells you the GPU is usable plus which model and driver it saw. Run it again after any driver upgrade or VFIO bind — that's the check that catches a broken host before a provisioning request does.
--gpu takes a device index (0) or a PCI address (0000:0b:00.0). The LXC backend wires the device in via Incus nvidia.runtime passthrough, so the container sees the real device rather than a shim.
The API field is a list of device IDs, not a single value, so a box can take one card, several, or all of them. Multi-GPU training and tensor-parallel inference need devices in the same box — they can't be spread across separate sandboxes — which is why this is a list and not a boolean.
The box runs an MCP server and is reachable over SSH, so Claude Code, Cursor, or your own agent drives it with no GPU-specific glue. From inside, it's just a Linux host with a GPU.
validate-gpu
probes and what we run. It is a self-hosted / bring-your-own-hardware
story: there is no GPU tier on the hosted cloud today. If you want GPU boxes managed for you,
tell us — that
demand is what would move it up the list.
FAQ
In a self-hosted box, yes — attach one or more NVIDIA GPUs at creation time and the agent inside sees them like any other host, nvidia-smi and CUDA included. Most managed agent sandboxes are CPU-only, so if the task needs a device, the environment has to be one you provision.
Pass --gpu at create time with a device index or PCI address: containarium create my-box --gpu 0. Repeat or comma-separate it for several.
Yes — the field is a list of device IDs, so a box can hold one, several, or all of them. That's required for multi-GPU training and tensor-parallel inference, which can't be split across separate sandboxes.
containarium backends validate-gpu launches a throwaway container, runs nvidia-smi, tears it down, and reports usability plus model and driver. It takes ~30s, and it's admin-only since it creates and deletes a container. Use --pci to check one specific card.
That's the intended path. The core is Apache 2.0 and installs on a machine you own, so a card you've already bought becomes agent capacity at close to zero marginal cost — no per-hour rental, and weights and data never leave your network.
Not today — GPU boxes are a self-hosted / bring-your-own-hardware story, and we'd rather say so than list a tier that doesn't exist. Email us if you want managed GPU capacity; that demand is what would prioritise it.
Install the open source on the machine the card is already in, and validate passthrough in about thirty seconds.