Run your own · your model

The agent on a model you run

Serve a model on a machine of yours, a card under the desk, the Mac or a box rented for the afternoon, and the agent can run on it. It shows in a scene's picker beside the models we host, its calls go through the machine's own connection, and no token is billed or sent to a model provider.

A compact GPU workstation tower on a wooden workbench, the graphics card visible through the glass side
What it means

Where you add the model's cell decides who can pick it

Run a model server on the machine (vLLM, SGLang, Ollama, MLX, anything that speaks the OpenAI API) and publish its port, as you would any service. Then add a cell to a scene, Model on a machine, naming the machine and the service. Every model the server lists shows in that scene's picker as qwen3.8-27b on lab-gpu, and you pick it the way you pick any model.

Calls go through the connection the node already holds open: no port on your router, no public address. A server that wants a key gets it from a scene secret the cell names, so the key is in no prompt and no code. Tokens are not billed, and every call is recorded in the trace like any other.

Add the cell to a scene and that scene can use the model; add it at the workspace level and every scene in it can. A cell in your private layer serves only your own runs, and a call from anywhere else is refused, never sent to another model instead.

Models in this sceneWHERE IT RUNS
qwen3.8-27b on lab-gpuyour box · 96 GB · nothing per tokenyours
gemma-4-12b on desk-minithe Mac · MLX · one at a timeyours
gpt-oss-120b on h100-afternoonrented · by the hourrented
Claude, GPT, Gemini …on your keys, then ourshosted
Each is named for the model and the machine serving it. A machine that is off lists nothing, so the picker only offers a model that can answer.
What fits where

Memory decides the model; the model decides the machine

MachineMemoryRuns well
Mac mini / Studio (Apple silicon)24–96 GB unified7B–70B with MLX, fast for one person at a time; not a serving box
A tower with an RTX 509032 GB27B-class at 8-bit, with room for a long context
A tower with an RTX PRO 600096 GB70B at FP8, GPT-OSS-120B, 27B at 262k context, several agents at once
Rented H100 / L40S80 / 48 GBAnything above, by the hour; gone when you leave
On-prem rack next to SAPwhat IT gives itThe gateway, the sandbox and the model on one box inside the walls
Does anything leave the building?
Prompts and answers pass through our server, which reaches your machine over the connection the node holds open. They never go to a model provider. To keep them inside entirely, run Daslab itself on your own hardware: self-hosting.
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. Every call is recorded in the trace either way.
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?
No. A scene's picker offers your models beside the hosted ones, and which runs is chosen per job. A job on your model stays on it: when the machine can't answer, the call fails rather than moving to a hosted model.