Skip to content
offrigoffrig

App and CLI

The app and the CLI do the same work over the same library. Both use the plain lane: one SSH alias offrig, the tunnel on 127.0.0.1:11435, and pods named offrig-<profile>. A side-car’s pods live in their own lanes and are never touched from here (see Lanes and job pods).

The app

Section titled “The app”

Start offrig-app. From one window it:

  1. shows your RunPod balance, live GPU prices and how long the balance lasts;
  2. launches the pod for the profile you pick, and pulls its models on the pod;
  3. opens the SSH tunnel;
  4. adds the pod’s models to Zed as their own provider;
  5. runs the seven guard checks;
  6. shuts the pod down, or terminates it after a stretch with every GPU idle.

Cancel launch stops a launch that is waiting for GPUs; nothing is rented while it waits. Closing the app with a pod running asks whether to terminate the pod or leave it running.

The CLI

Section titled “The CLI”
Command What it does Costs money
offrig status Balance, spend per hour, runway, and every pod on the account no
offrig gpus --count 2 [--min-vram 160] Live secure-cloud offers for a GPU count, cheapest first no
offrig profiles The tiers in your config, with their models and sizes no
offrig init Write the default config file if there is none, and print its path no
offrig up <profile> Launch, pull the models, wire Zed, run the checks, hold the tunnel yes
offrig up <profile> --wait 180 Wait up to 180 minutes for GPUs, renting nothing meanwhile yes, once they free up
offrig up <profile> --detach Set up, then exit instead of holding the tunnel yes
offrig up <profile> --no-zed Set up and do not write Zed’s settings yes
offrig tunnel <profile> Hold the tunnel to a running pod, with idle auto-stop the pod’s rate
offrig pull <model> [<profile>] Pull another model onto the pod; the pull runs on the pod the pod’s rate
offrig models [<profile>] Models on the pod, read over SSH no
offrig check [<model>] A streamed chat with a tool call through the tunnel, the way Zed sends it the pod’s rate
offrig guard [<profile>] Run the seven “never on my GPU” checks no
offrig zed [<profile>] / offrig zed-remove Write or remove offrig’s provider in Zed’s settings no
offrig connect [/workspace] Open the pod in Zed for remote editing the pod’s rate
offrig down <profile> --yes Terminate the pod; its disk goes with it stops billing
offrig budget [--provider runpod|openrouter] [<usd>] Set this project’s caps (human only): one per provider, plus an optional overall ceiling. With no amount in a terminal, a menu; --show prints every cap beside each provider’s own balance. Never below spent + committed no
offrig index <paths...> [--status] [--rebuild] Index this project’s files for hybrid search. Embeddings come from the CPU-only embed server, EmbeddingGemma 2 by default (embed_url, default http://127.0.0.1:11490). --status reports and changes nothing. --rebuild embeds every chunk again no
offrig verify calibrate <gold.jsonl>... Score a local loopback model as a verifier. Cloud model names are refused. Flags are in Reference local compute only
offrig stage <profile> --dc <DC> [--yes] Stage a recipe profile’s weights on a network volume yes, monthly

offrig up refuses when your runway with the pod running would be under one hour, because at zero RunPod stops every pod on the account. --yes overrides that refusal.

Zed

Section titled “Zed”

The pod’s models are a separate provider called offrig, pointing at the tunnel. If the pod is down, choosing one of its models errors; Zed never falls back to another provider. offrig up <profile> --default-model <model> also makes one of them Zed’s default agent model. offrig zed-remove takes the provider out and restores the previous default.

The provider is deliberately not named runpod. Zed would then read RUNPOD_API_KEY and send it to the model server.

Money safety

Section titled “Money safety”
  • Before a launch, offrig shows the cheapest free match and your runway with the pod running.
  • offrig up and the app do not start a watchdog. If the command stops with an error (a hang, an unresponsive engine, an engine that exited, or a timeout), the pod keeps billing until offrig down <profile> --yes or the app’s Shut down. A pod RunPod reports as ERROR or EXITED is terminated during the launch.
  • Auto-stop terminates the pod after 30 minutes with every GPU under 5% busy. The minutes are set by auto_stop_idle_minutes in the config; auto_stop_idle_minutes = "off" turns it off.
  • offrig only changes pods it named. Every other pod on the account is listed and left alone.

The seven guard checks

Section titled “The seven guard checks”
Check Fails when
Tunnel avoids the local Ollama port the tunnel port is 11434
Zed sends pod models through the tunnel Zed’s provider URL is anything but the tunnel
The pod’s Ollama is not exposed to the internet the pod maps port 11434 publicly
The tunnel ends at the pod the model list through the tunnel differs from the list read on the pod over SSH
Pod models are not on this machine a pod model also exists in the local Ollama
No pod model shares a name with a local Zed model a name in the offrig provider is also in Zed’s local Ollama list
Every model Zed offers is on the pod Zed offers a model the pod does not have