> ## Documentation Index
> Fetch the complete documentation index at: https://comfyui-mcp.artokun.io/docs/llms.txt
> Use this file to discover all available pages before exploring further.

# July 14: the self-updating orchestrator, vision for every backend

> The panel orchestrator now updates and restarts itself when every agent is idle — a running install never goes stale. Every Ollama-family backend and ChatGPT-direct gains inline image delivery with an honest fallback when a model is text-only, the gemma4 fine-tunes get correct sampling, and a stable phone-pairing token survives restarts.

*July 14, 2026*

v0.31.1 → v0.34.0 plus panel 0.8.1 / 0.8.2. Two headliners:
[the orchestrator that never goes stale](../blog/self-updating-orchestrator) and
[giving every backend eyes](../blog/vision-per-model).

## Self-updater + self-restarter

The panel orchestrator (default ON) re-checks npm hourly, updates the installed
package, and **restarts itself once every agent is idle** with nothing queued, held,
or rendering — the panel announces the restart, sessions resume from the durable
store, and the panel reconnects on its own. Dev installs are never modified on disk;
instead the orchestrator restarts when a rebuilt `dist` lands, so `npm run build` is
all a developer needs. MCP stdio mode never self-restarts (the client owns that
process). Opt out with `COMFYUI_MCP_AUTO_UPDATE_DISABLE=1`. And
`npx github:artokun/comfyui-mcp` now works as a nightly channel.

## Vision is per-model, for every backend

Inline image delivery lands for **every Ollama-family backend** (ollama, OpenRouter,
LM Studio, llama.cpp, custom endpoints, GLM, Kimi, Copilot) and for **ChatGPT-direct**
via Responses-API `input_image` data URLs. Vision is always attempted; when an
endpoint rejects image input there's a graceful one-shot strip-and-retry plus an
honest 📎 note — and render-completed events no longer tell a text-only model the
image is "attached below" (confabulation guard).

The **gemma4 fine-tunes also got their sampling fixed**: the tags' baked
`temperature 0` caused greedy repetition loops; the backend now sends the
Gemma-recommended sampling (temp 1.0 / top\_k 64 / top\_p 0.95) for the fine-tune tags.

## Mobile: a pairing token that survives restarts

Set `COMFYUI_MCP_PAIR_TOKEN` to pin the mobile bridge's pairing token (#219) — a
paired phone reconnects across orchestrator restarts instead of dying on a
per-session token, and the LAN listener auto-starts at boot with a ready-to-paste
pairing URL. Unset keeps the previous on-demand, nothing-exposed-until-you-ask
behavior.

## Also shipped

* **Strip/slice read the live canvas** — `panel_strip_workflow` /
  `panel_slice_workflow` called with no source capture the open graph directly (new
  `graph_serialize` command, panel 0.8.2) — no more save-to-disk round trip.
* **Keys live in one place** — `~/.comfyui-mcp/.env` is the only dotenv location (a
  legacy package-root `.env` auto-migrates once); panel users manage keys via the
  API Keys card. Boot logs which providers have a key and its source — never values.
* **Connection guide for the five new providers** (#214) — Grok / Kimi / GLM /
  ChatGPT-direct / Copilot sign-in paths and picker roster.

## Fixes

A user interrupt (Stop or "Send now") no longer paints a false "turn failed" banner
(#221); UI→API conversion no longer scrambles widget values on nodes with custom
serialized-widget layouts like LTXDirector (#222); the tab-id migration self-heal
was hardened (#212, #213); the model catalog is live and account-aware, with current
DeepSeek in the curated picks; and the expired Discord invite was replaced with the
permanent link (#220).