Rixi
Give your agents a sealed machine of their own
Run any project remotely as if it were local, and hand agents a throwaway box you control.
Kevin Keller · kevinkeller.org
Why it matters
Agents need somewhere to run code, and both usual options leak
- Your laptop: the agent can touch your files, credentials and network.
- A hosted sandbox: your code, data and model calls go through someone else's platform.
- Rixi's answer is a box: a machine you control, created with one command and destroyed when done.
The core
Ship the code and its exact environment, then stream the output back
- Package the directory, code and its resolved Pixi environment.
- Upload it over an encrypted channel.
- Run the task on the remote machine in an isolated subprocess.
- Stream the output back.
No Dockerfile, no image registry, no "works on my machine".
One command
From a cloud API key to a ready box, and back to zero
rixi upcreates the server, installs Rixi, negotiates a key and saves a profile.- Hetzner, Scaleway or AWS (experimental), CPU or GPU sizes.
- Pinned TLS per box and a fresh two-minute token per request.
rixi downdestroys it and billing stops.
The key idea
Seal the box from the outside
Egress is blocked or allowlisted in the cloud provider's firewall, where nothing running on the box as root can undo it.
Model access
The riskiest call an agent makes becomes an explicit choice
| Mode | What happens |
|---|---|
governed | Through a gateway and policy engine: PII rules, signed audit trail, your provider key never enters the box |
direct | Your own endpoint with your own key, nothing in between |
none | The default: no model access at all |
Agents in the loop
Boxes become tools for any MCP client
rixi mcplets an agent create a box, run commands, move files and destroy it.- It runs on your machine, so output and files never pass through a hosted service.
- It destroys only the boxes it created itself.
- Hosted boxes shut down after a time-to-live, two hours by default.
Hard problems, one primitive
Ship a project and its environment, and many problems disappear
| Problem | Rixi |
|---|---|
| Environment won't reproduce on the GPU box | Ships the exact resolved .pixi/ |
| Box is air-gapped | --offline-mode bundles the environment |
| Box is behind a firewall | Encrypted reverse tunnel, no inbound ports |
| One Metaflow step needs a GPU | Add @rixi to that step |
Security model
Security is not an add-on when you run code remotely by design
- Per-request tokens: ES256, two-minute expiry, one box
- Pinned TLS plus end-to-end AES-256-GCM
- Loopback by default: no auth, no start
- Hardened uploads: credential-looking files left out
- Pinned installs: checked against recorded hashes
- Optional unprivileged mode: agent code without root
Scaling up
A gateway turns single boxes into a governed fleet
- On-demand provisioning with OpenTofu on Hetzner, Scaleway or Kubernetes.
- Teardown on idle, time-to-live or release, so billing stops.
- Cheapest-box scheduling and budget caps on hourly burn and total spend.
- Admin policy floor, role-based access and a queryable audit trail.
Where it fits
Strongest where environments must reproduce and machines are locked down
- Exact environments on GPU or remote boxes.
- Firewalled or air-gapped machines.
- Agents that need a sealed machine you control.
Not a hosted-only platform and not a workflow orchestrator: pair it with your scheduler or Metaflow.
Rixi
Let agents work on a real machine, without handing them your laptop.
Kevin Keller · Personal open-source project (MIT), not an official product