Loading the catalogue…
Stav Tunnel
Put your own hardware behind your Stav key
Plug the Mac mini or the GPU server you already run into your Stav workspace. The models stay private to you. Your apps keep one base URL, one key and one request log.
Your appapi.stav.aione key
Private provider
Your hardware
mac-mini-01llama-3.3-70b
mac-mini-02nb-llama-70b
gpu-rack-01qwen3-235b
Public providers
BergetSE
MistralFR
RegoloIT
AnthropicUS
Your hardware is a private provider in your catalogue, next to the public ones. Any location, same key, same Routers, same request log.
What people put behind the tunnel
Everything from a Mac mini on a desk to a dedicated GPU server in your own rack. One Stav Tunnel daemon per machine, one workspace for all of them.
Consolidate your Mac minis
Three Mac minis under three desks become one catalogue. Each is a node; the Router picks whichever is online.
Give your team shared access to local hardware
One GPU box, every developer. Each gets a Stav key with a spend cap, and every request lands in the log. Nobody needs SSH.
From Mac mini to dedicated GPU server
Start with the machine you have. Add the server later under the same model name; nothing in your app changes.
Serve the model the catalogue does not have
A fine-tune, a Norwegian Whisper, a quantised build. Private to your workspace, or mapped to its catalogue page.
No network setup. Nothing to open
The usual way to reach a machine behind a firewall is to punch a hole in it. Stav Tunnel never does. The machine reaches out instead.
Your machine dials out
The Stav Tunnel daemon opens one connection from the machine to Stav, the same direction a browser uses to open a web page. Every firewall already allows that.
Requests ride back down that line
Your app calls api.stav.ai as usual. Stav hands the request down the open connection, your machine answers, and the answer comes back the same way.
Close it, and it is gone
Stop the Stav Tunnel daemon, or revoke the node in your workspace, and there is no path left. Nothing stays open.
Things you will not do
- Open a port
- Write a firewall rule
- Set up a VPN
- Get a static IP address
- Install a certificate
- Book a change window with IT
What you will do: run one command on the machine. Setup takes as long as installing a package.
Works with the servers you already run
Start the server the way you do today. The Stav Tunnel daemon finds it on the machine and reports the models it serves. Nothing to reconfigure.
OllamaThe easy way to run a model on a Mac or a laptop.
LM StudioA desktop app with a built-in server. Switch it on in the app.
llama.cppThe lightweight server that runs on CPUs and small GPUs.
vLLMThe common choice for GPU servers. Chat, embeddings and Whisper.
SGLangA GPU server built for many requests at once.
Running something else? Any server that speaks the OpenAI chat API works when you point the Stav Tunnel daemon at it.
Install Stav Tunnel
Replace <your-token> with the token from your workspace.
macOS or Linux with Homebrew. The last command makes it a background service, so it runs at login and after a restart.
brew install stavhq/tap/stav-tunnel stav-tunnel up --token <your-token> stav-tunnel service install
brew install, and it is in your catalogue
Add a node in Connect → Stav Tunnel to get your token. One command on the machine, and the models are in your workspace.
One request log, your nodes included
A request to your own machine lands in the same log as a request to a public provider. Same columns, same filters, same export. The only difference is the cost.
Nothing to wire up
The log is written at the gateway, so it already knows your nodes.
Cost reads €0.00
Tokens are still counted, so usage per model and per key still adds up.
Open any row
See why it went there: the Router’s choice and the runners-up, for your node as for any other.
Request logProvider: all · Last 24 h
TimeModelProviderCountryCost
14:02:11nb-llama-70bYour hardware · mac-mini-02NO€0.00
14:01:57mistral-largeMistralFR€0.0041
14:01:40llama-3.3-70bYour hardware · mac-mini-01NO€0.00
14:01:22qwen3-235bYour hardware · gpu-rack-01NO€0.00
14:00:58gpt-oss-120bBergetSE€0.0007
14:00:31nb-whisper-largeYour hardware · gpu-rack-01NO€0.00
Every Router has a Tunnel nodes setting
One field in the Router editor decides how your machines take part. Set it per Router, so a test Router can stay on the catalogue while production prefers your nodes.
Prefer
An online node goes first. The Router falls back to the catalogue when it is away.
Only
Your nodes alone. Nothing online means the request fails, never a quiet fallback to a third party.
Ignore
The Router stays on the catalogue. Call a node by name when you want it.
Connect → Routers → @acme/productionPublished
Tunnel nodes
Preferyour nodes first, then the catalogue
Onlyyour nodes, fails closed
Ignorecatalogue only
CancelSave changes
Add your hardware to Stav’s gateway
Add a node in Connect → Stav Tunnel. The first model shows up before the coffee is done.