offrig Handbook

offrig runs big models on rented RunPod GPUs, with a guarantee: they never run on your own GPU. It is three front ends over one Rust library:

• offrig-app, a Windows desktop app: pick a tier, press Launch pod, and the pod’s models appear in Zed.
• offrig, a CLI for the same work from a terminal.
• offrig-mcp, an MCP side-car that lets an agent such as Claude Code plan a paid session under a human-set budget, rent GPUs, run work on them, and shut them down.

What it is for

Models too big for your machine (a 120B model on one 96 GB card, or a 480B model across four), and GPU work your machine shouldn’t do: training runs, renders, evaluations. offrig rents the hardware by the hour, gets it ready, and keeps the bill bounded.

The three promises

1. Nothing runs on your GPU. The model server is reachable only through an SSH tunnel. The tunnel refuses your local Ollama’s port, so a dead tunnel fails instead of falling through to the local server. Seven checks verify this on every launch.
2. No pod outlives its plan. Every side-car launch starts a watchdog: a separate process that terminates the pod at the plan’s deadline, even if everything else is gone. A setup that fails before the engine is serving terminates the pod. A recipe that has started and then hangs, or that cannot answer one completion, is reported and left running until you shut it down or the deadline ends it.
3. An agent cannot name its own price. A human sets each project’s budget cap. A plan commits its worst case against it before anything is rented, and a launch takes only a plan id.

Where to go next

If you want to	Read
Install offrig and run a first pod	Getting started
Use the app or the CLI day to day	App and CLI
Give an agent a GPU and a budget	The side-car
Change tiers, GPUs, prices or disks	Configuration
Run several projects at once, or GPU jobs	Lanes and job pods
Know exactly what offrig touches	Security and what it changes
Look up a command, tool, error code or exit code	Reference

Status

offrig has run real work since 2026-10-02. Highlights:

• A frontier run on 4 × RTX PRO 6000 served Qwen3-Coder-480B on SGLang and drained 31 handoffs in 20 seconds, for $3.59.
• Training runs for aspire-si on job pods.
• Singing renders for ai-jam-sessions on cheap A40 pods.
• Two projects using it at once, each in its own lane.

The local GPU stayed idle throughout.

On 2026-10-11 a SGLang pod logged that startup had finished, then /health stayed down for 7 minutes while /v1/models still answered, on a CUDA 13.2 host. Working pods that week were on 13.0. 13.2 is not shown to be the cause; the pin lets a profile stay on the version that worked. The ready check and the exact CUDA pin in this handbook describe what 1.1.0 does about that. They have not been re-run on a rented pod. Tests cover them with a fake ssh session and a fake RunPod.