The agent on a model you run
Your keys already serve before ours. The next rung is the model itself: one you serve on a card under the desk, on the Mac, on a box rented for the afternoon, or on the rack next to SAP. It appears in the picker beside the ones we host, the agent uses it like any other, and nothing — not a token, not a file — leaves the building.
One more model in the picker, with a different price and a different address
A machine that serves a model — vLLM, SGLang, Ollama, MLX, anything that speaks the OpenAI API — puts it on its tile, and next into the model picker of every scene in the workspace as lab-gpu / Qwen3.8-27B. Pick it for a scene, or name it in a prompt, the way you pick any model today.
The call travels through the connection the node already holds open: no port on your router, no key to hand out, no public address. The price is what the machine costs you — nothing per token — and usage is still recorded, so a scene's trace reads the same whether it ran on our model or yours.
Memory decides the model; the model decides the machine
Where the model runs is a machine you add
One command, a code on your phone, vllm serve. The card under the desk is the cheapest tokens you will ever buy.
Apple silicon runs a 12B–70B model with MLX. Good for one person and a quiet evening; not for a team.
Add a Mac mini →No card? Rent an H100 by the hour. It pairs like any machine and is gone when you leave.
Rent one →The gateway box that already mirrors SAP can serve the model too. One binary, inside the walls.
The gateway →- Does anything leave the building?
- No. Prompts, files and answers travel between our server and your machine over the one connection the node holds open, and never to a model provider. If that is still too much, the gateway puts our server inside the walls too.
- What does it cost?
- The machine. Tokens on your own model are not billed; a rented box costs what the provider charges by the hour. Usage is recorded either way, so a scene's trace is complete.
- Which model should I serve?
- For agents: Qwen3.8 27B on 24–32 GB, GPT-OSS-120B or a 70B at FP8 on 96 GB. Vision and policy models for the bench are a separate question the GPU box guide answers.
- Can the model be the only one a workspace may use?
- Next: a workspace setting that pins the picker to your machines, so no scene can reach a hosted model. Today the picker offers both and the choice is per scene.