July 21: rent a GPU from the panel — RunPod connector and cloud LoRA training

A first-party RunPod connector deploys and drives a cloud GPU pod from the panel and phone with an honest local⇄pod host switch, live cost/GPU status, and idle auto-stop; the LoRA trainer learns to train on rented pods; plus resolve_missing_models and provider model discoverability.

July 21, 2026

v0.44.0 plus the unreleased 0.45.0 line, panel, and mobile — the week ComfyUI went to the cloud without leaving your custom canvas behind.

Rent a GPU without leaving the panel

The first-party RunPod connector lands (v0.44.0): runpod_pod_create one-tap-deploys the project’s own RunPod template through the referral link, falling back across GPU types and COMMUNITY→SECURE cloud when capacity is tight, and the existing status/start/stop/troubleshoot/connect tools manage any live pod by ID. The pod runs the full comfyui-mcp environment, so the agent installs your custom nodes + LoRAs and downloads your models onto it — exact canvas parity, not a stripped-down managed template.

An honest host switch ties it together: a comfyui_target broadcast pushed on every retarget drives a panel host pill that reads 🟢 Local · your rig or **🔵 RunPod · <pod> · GPU · $/hr**, so where a render runs is never ambiguous, and `runpod_use_local` flips between them in one tap. A `runpod_status` broadcast feeds a live card (GPU / VRAM / uptime /$·hr), and idle auto-stop shuts an idle pod down for you — but only while you’re actually rendering on it, so it never yanks a pod mid-session (RUNPOD_IDLE_STOP_MINUTES, default 15). RunPod serves ComfyUI on port 3000 by convention, and the connector targets 3000 for you. The full walk-through.

Train a LoRA on a rented GPU

The CLI LoRA trainer gains a dockerless native driver and an SSH transport that stands ai-toolkit up on a RunPod pod (#263): an idempotent bootstrap (clone@pin → venv → torch cu128 → requirements), a dataset rsynced up, live step/loss progress streamed back over ssh, and stop/liveness via pkill/pgrep — all behind the same seven train_* tools and the same crash-safe job registry as local training. FLUX.1-dev LoRA training no longer needs a 24 GB card you own; borrow one for the hour the run takes. The story.

Also new

• resolve_missing_models (#267) — one call finds every model a workflow needs but the server doesn’t have, and proposes VRAM-aware download candidates (size, source, precision/quant, and a fits/too-big verdict against real /system_stats VRAM). Detection is mapping-free — a model-looking value absent from its own ComfyUI combo is missing — so it covers checkpoints, LoRAs, VAEs, ControlNets, UNets, CLIP and custom-pack types alike. “Open this Template and make it runnable” is now one hop, not two.
• Provider model discoverability (#264) — the api-key credential card now says which model a provider is actually on (env override if set, else the pinned default) and names the env var to change it, generated from the registry so it can’t drift. Answers “why am I not on the model I set?” for GLM / Kimi / Moonshot.
• Antigravity (agy) backend (#262) for Google-subscription users, hardened this week: no secrets at rest, ownership-aware config lifecycle, turn-settlement guarantees, and --effort, verified live against agy 1.1.5 (#271).

Mobile

• RunPod control sheet + honest host indicator (#22) — the phone gets the same cloud control surface as the panel: a host banner, a live status card, a pod dropdown by name, Connect / Start / Stop / Use-Local, and a confirm-armed Deploy.
• One-tap Diagnose on a failed render (#21) — the queue bar notices when the run it was watching failed and offers a Diagnose action that root-causes it via the read-only diagnose_run tool, instead of leaving you to copy the error back into chat.

Fixes

• close a live secret-leak + billing bug path on RunPod pod control (#270)
• prefer HF_TOKEN over HUGGINGFACE_TOKEN for the Hugging Face token
• close deferred RunPod / training / review findings (#268, #269, #273, #274, #276, #277)