NexusRun — portable AI units
See it run<br>Not a chatbot. An agent that does the work.
Here it is answering a question by going and finding the answer — on a<br>laptop, with no network and no API key. Real output, not a mockup.
terminal<br>100% local
nexus run notes-agent -p "When does the launch window open?"
running notes-agent:1.0.0 on CPU via llama.cpp/server<br>starting MCP server notes<br>notes v1.0.0 — 2 tool(s): notes__list_notes, notes__read_note
→ notes__list_notes({})<br>✓ notes__list_notes — launch.txt sites.txt<br>→ notes__read_note({"name":"launch.txt"})<br>✓ notes__read_note — The launch window opens on 14 March 2027.
The launch window opens on 14 March 2027.
— 54 tokens · 7.0 tok/s · 2 tool call(s) over 3 turns
# nexus.yaml — the whole agent<br>apiVersion: nexusrun.dev/v1<br>name: notes-agent<br>version: 1.0.0
models:<br># already pulled with Ollama? reuse it<br>- source: ollama:llama3.1:8b
entrypoint:<br>type: chat<br>system_prompt: |<br>Answer using the note tools.
mcp_servers:<br>notes:<br># fetched, pinned and sandboxed for you<br>source: github:acme/notes-mcp#v1.0.0<br>command: ["node", "server.js"]<br>sandbox:<br>allowed_paths: [/home/me/notes]
secrets:<br># declared here, never stored here<br>- name: API_KEY<br>required: true
One readable file, not a Dockerfile
Everything that decides how the agent behaves is in front of you. It<br>points at the model rather than containing it, so the file stays about<br>a kilobyte instead of five gigabytes.
Its tools come with it
Tool servers are dependencies, pinned like any other. Nobody has to<br>install anything first, and each one is confined by the kernel to the<br>paths you listed — undeclared means denied.
Secrets stay out of the artifact
The file says which key it needs. The value lives encrypted on the<br>machine that runs it, so the agent is safe to commit and safe to share.
Why it's different<br>It's built for machines you don't control.
Someone else's laptop, a rack of mixed hardware, a kiosk in a shop, a box<br>that has never seen the internet. That is where agents actually have to<br>work, and where the usual assumptions quietly fail.
It runs where your data already is
On the machine that holds the records, behind the VLAN, or on a box with<br>no network at all. Seal the model in and carry one file across. Nothing<br>is uploaded, and running it more often costs nothing extra.
Verified: imported to an empty machine and generated with no network at all.
It won't quietly use the wrong one
A good GPU the installed runtime was never built to drive is the normal<br>case, and most tools fall back to the CPU without a word. This one says<br>so — then scores the agent on that machine, because answers change with<br>the hardware and the model, not just the prompt.
Verified: a model that scored 0/3 there was rejected and a better one chosen, automatically, per machine.
Safe to run someone else's agent
Read the file, see exactly what it will reach for, then run it knowing<br>the kernel holds it to that. Its tools, its files, its network — nothing<br>it did not declare is possible.
Verified from inside the sandbox: no stray file writes, no outbound connections.
NexusRun doesn't replace llama.cpp or Ollama — it runs on top of them. What<br>it adds is the packaging, the sharing, and the straight answers.
Get started<br>Three commands.
Create it
nexus init my-agent --model llama3.1:8b
Writes one small YAML file. Edit the prompt to make it yours.
Run it
nexus run ./my-agent -p "hello"
Straight from the folder, so you can change the prompt and go again.
Send it anywhere
nexus push my-agent:0.1.0 ghcr.io/you/my-agent:0.1.0
A few hundred bytes, to any registry you already use. No network at the<br>other end? nexus build --seal packs the<br>model in so it travels on a USB stick.
Hosting for your team →
The runtime is free forever. NexusRun Cloud hosts and signs<br>your units for teams shipping to real fleets —<br>early access is open.