Push your agents to a private registry, pull them onto any machine you run, and carry them in by file where there is no network at all.
Inference stays on your hardware. We never see your prompts, your data, or your tokens.
Free for 90 days. No card required.
The trouble starts when the agent has to run on 40 machines you don't administer: a locked-down Linux box at a clinic, a line PC on a plant floor, a kiosk, a host that never touches the internet.
Every machine pulls the same versioned unit from one place, instead of a shared drive and a hope.
Your units live in your own namespace. Nobody else can list them, and pulls need a token you issued.
Export a sealed file, carry it across, and import it. The registry is optional on the far side.
Public registries aren't private, Kubernetes is too much for a kiosk, and Docker often isn't allowed on the box in the first place.
A unit is the model reference, system prompt, MCP tool servers, secret declarations, and hardware preferences. It is usually under 1 KB because it references the model instead of embedding it.
export NEXUSRUN_REGISTRY_USER=you
export NEXUSRUN_REGISTRY_PASSWORD=nxr_pat_… # your access token
nexus build ./triage-agent
nexus push triage-agent:1.4.0 registry.nexusrun.dev/acme/triage-agent:1.4.0
It's a standard OCI registry, so nexus push is all it takes.
nexus pull registry.nexusrun.dev/acme/triage-agent:1.4.0
nexus doctor # reports what the hardware can actually run
nexus run triage-agent:1.4.0 -p "summarize today's intake"
Downloads come from a global CDN, so a whole fleet pulling the same model fetches it from our storage once.
nexus build --seal -o triage.nx # embeds the weights nexus export --offline # checks the package is complete nexus import triage.nx # on the machine with no network
Move the file however you move files today: USB, data diode, approved transfer.
| Stays on your machines | NexusRun Cloud hosts |
|---|---|
| Model weights, unless you choose to push a sealed unit | Unit artifacts (OCI) you push |
| Prompts, inputs, outputs, logs | Your org, its members and access tokens (stored hashed) |
| Secret values, in a local encrypted store | Storage and bandwidth usage for billing |
| Inference: llama.cpp or Ollama on your hardware | |
| The NexusRun runtime (Apache 2.0, free forever) |
There is no hosted inference and no per-token bill. If NexusRun Cloud went down tomorrow, units already on your machines would keep running.
Pay for the registry, not for running agents. Running agents is never billed.
| Free trial | Developer | Team | Enterprise | |
|---|---|---|---|---|
| Price | $0 for 90 days, no card | $19/mo | $149/mo | Custom |
| Storage included | 5 GB | 10 GB | 100 GB | Negotiated |
| Pull bandwidth / mo | 20 GB | 100 GB | 500 GB | Committed |
| Seats | 1 | 2 | 5 | Custom |
| Private repositories | ✓ | ✓ | ✓ | ✓ |
| Immutable tags | ✓ | ✓ | ✓ | ✓ |
| Support | Community | Community + email | Priority email | SLA |
Pulls beyond the included bandwidth are billed at $0.09/GB. Storage is capped at the included amount, and upgrading raises it. The trial stops at its limit rather than billing you.
The NexusRun CLI and runtime are Apache 2.0 and free forever. Cloud is optional: you can always push to any OCI registry you already run.
Yes. The nexus CLI and runtime are Apache 2.0 and stay free, with no host limits. Cloud charges for registry storage and bandwidth only. Running agents is never billed.
We never see prompts, inputs, outputs, or secret values. Inference runs on your hardware. We store only the unit artifacts you push; a typical unit is metadata plus a model reference, under 1 KB. If you push a sealed unit, the weights are stored as OCI layers in your private namespace. To keep weights off our servers, reference the model and distribute it yourself.
The registry recomputes the SHA-256 of every blob it receives and rejects anything that doesn't match the digest you claimed. Access tokens are stored only as hashes, and the short-lived tokens that authorize each pull or push expire in five minutes.
Yes. Build a sealed .nx file, move it across however you move files today, and nexus import it. For clean Linux hosts with no inference engine installed, nexus bundle also packages the NexusRun binary and a pinned llama.cpp runtime.
No. NexusRun is one static binary, about 11 MB. There's no container runtime and no daemon you have to install. Units run under a Landlock sandbox on Linux.
Today: Linux on x86-64, CPU inference. nexus doctor tells you what a given machine can actually use, not just what it detects.
Yes, any time. nexus push works with any OCI registry (GHCR, Harbor, ECR, and others), and anything you store with us can be pulled back out with standard tools.
Both package agents as OCI artifacts. cagent needs Docker on the target. NexusRun doesn't, which is what matters on locked-down, embedded, or air-gapped machines.
Push a unit, then pull it onto a machine with networking turned off.
Fleet or air-gap rollout? Email hello@nexusrun.dev for a 20-minute demo.