Skip to main content
by artokun · July 16, 2026 · mobile · pairing · remote control The render finished while you were making coffee. That sentence is the whole problem with ComfyUI, and it always has been. A good generation — a video especially — takes long enough that you leave. You queue the job, you wander off, and now the state of your work lives on a screen in another room. Did it finish? Did it OOM at 90%? Did the agent install the missing node and retry, and is it now waiting on a question you can’t see? The desktop is doing fine. You’re the disconnected part. The ComfyUI-MCP mobile app closes that loop. It’s a phone-native front end — Flutter, shipping on iOS and Android in closed beta — to the same agent that powers the Sidebar Panel. Not a dashboard, not a read-only queue viewer: the actual agent, on the actual bridge, with your desktop’s GPU doing the work.

​
What it actually is

The app connects to the panel orchestrator’s bridge — the same WebSocket the in-ComfyUI sidebar rides — so it’s not a parallel system with its own limits. It’s another seat at the same table:
  • The same conversation. Chat with the agent from the couch exactly like you would in the sidebar. Ask for a render, ask it to fix a workflow, ask why the last one came out wrong.
  • Renders land in your hand. Finished images appear inline; videos play inline with a lightbox, and you can save MP4 outputs straight to your camera roll. A live queue monitor bar shows the running job, queue depth, and sampler progress — powered by the orchestrator’s change-only queue_status broadcast — with a one-tap cancel that’s deliberately narrow: it only interrupts the exact prompt_id you’re looking at, and never touches pending jobs (which might belong to another agent sharing the queue).
  • Full CivitAI browsing — real parity, not a teaser. Search models, page through all results with infinite scroll, browse creators (the creator leaderboard and username search ride the same download_model action:"search_creators" tooling the desktop got in v0.35.0), and see your favorites properly — the app reads your CivitAI likes collection, so if you’ve liked 700 models, you see 700 models, not the handful the reactions endpoint returns.
  • The same provider chips. The model picker exposes provider, model, and reasoning effort — Claude, ChatGPT, Gemini, Kimi K3, and the whole local/self-hosted family (Ollama, LM Studio, llama.cpp, OpenRouter). Whatever the orchestrator can run, the phone can select.
  • A LoRA hub and prompt presets. Browse installed LoRAs with detail sheets and an update checker (the civitai: provenance line from v0.37.0 is what lets it link back and check for newer versions), and keep favorite prompts one tap away.
The mental model is simple: the orchestrator is the brain, ComfyUI is the hands, and the phone is just another pair of eyes and a mouth. Nothing was forked to make mobile happen.

​
Pairing: one QR, every path

Pairing flows are where companion apps usually die — type this IP, copy this token, install this cert. Here it’s one QR code shown by the panel, and it’s deliberately multi-purpose:
  • Scan it from inside the app and you’re logged in instantly. Done.
  • Scan it with your phone’s camera — before you’ve even installed the app — and it opens a lander that reads your user-agent and routes you to the right place: TestFlight on iOS, Firebase App Distribution on Android. If the app is already installed, the lander’s Open in app button fires the comfyui-mcp://pair deep link and drops you straight into a paired session.
The security detail that makes this comfortable: the pairing credentials — host and token — ride the URL fragment, not the query string. Fragments never leave the browser; they’re not sent in the HTTP request, so your bridge token never appears in a server log, a proxy trace, or an analytics pixel. Scan the same QR three different ways and the secret still only ever exists on your two devices.

​
Pairing that survives restarts

Early beta testers found the sharp edge fast: the bridge’s pairing token was per-session. Restart the orchestrator — an update, a reboot, the self-updater doing its job — and every paired phone silently died and needed a re-scan. Fine for a demo, hostile for a daily driver. v0.33.0 fixed this with a stable pair token. Set COMFYUI_MCP_PAIR_TOKEN in ~/.comfyui-mcp/.env and the token is pinned: a paired phone reconnects across orchestrator restarts like nothing happened. When set, the LAN pairing listener also auto-starts at boot and prints a ready-to-paste ws://<lan-ip>:<port>/?token=… URL. The default posture is unchanged and worth stating plainly: if you don’t set the token, the pairing listener stays on-demand and per-session — nothing is exposed on your LAN until you explicitly ask for the QR. Persistence is opt-in; “nothing listening unless you asked” is the default.

​
The headline: mirror a desktop tab and drive it

Everything above makes the phone a good second client. Tab mirror (v0.35.0) makes it a remote control. The scenario: you’ve been working in the panel on your desktop all afternoon. Mid-conversation, mid-workflow, the render is queued — and you have to leave the desk. On the phone, you open remote control, pick that desktop tab, and the phone mirrors its live session: the same conversation you were having at the desk, streaming in real time. Type a message on the phone and it appears in the desktop tab — sent as that session, answered in that session, visible on both screens at once. You’re not starting a parallel chat with the same agent; you are operating the desktop tab from your pocket. That’s the feature where sloppy engineering would be genuinely dangerous — a remote client injecting into another client’s session is one bad assumption away from a cross-session leak. So the implementation is built on deliberate invariants rather than trust:
  • Server-authoritative tab identity. The phone never gets to claim which tab it’s mirroring; the orchestrator owns the canonical tab identity and the mapping. A client asserting “I am tab X” is ignored — identity is assigned, not requested.
  • Allowlisted fan-out. Mirrored frames are fanned out only to sockets explicitly allowlisted for that tab, keyed by canonical id. There is no broadcast path where a frame for one session could reach a socket that wasn’t granted it.
  • Pinned socket kind. A socket’s role is pinned at handshake. A pairing socket can’t quietly become a mirror socket mid-connection; each kind gets exactly the frames its role allows.
One honest limitation in the current beta: the mirror streams the session live from the moment you attach — full transcript hydration (scrolling back through everything said before you picked up the phone) is deferred to v2, because backfilling a long desktop conversation onto a phone screen is a UX decision we’d rather get right than rush.

​
Where the beta is honest

This is a closed beta, and the framing is deliberate. Builds ship through TestFlight on iOS and Firebase App Distribution on Android — beta channels, not the App Store and Play Store. Store releases come at launch; right now the release cadence is fast (the changelog is a steady drip of “favorites now show all 700, not 12”-grade fixes), and beta channels are the right place for that pace. What you get in exchange for tolerating the occasional rough edge is a direct line: mobile beta feedback on Discord has repeatedly turned into shipped tools within days — creator search existed on the phone because a beta tester asked for it. The through-line of the whole project has been removing the reasons you have to sit at the machine: the agent installs the nodes, recovers from the crashes, judges its own renders. The phone removes the last reason. The render finishes while you’re making coffee — and now the render, the agent that made it, and the conversation about what to do next are all in your pocket, cup in the other hand.
Put your ComfyUI agent in your pocket: install comfyui-mcp with the Agent Panel, then join the beta — the iOS TestFlight and Android Firebase App Distribution invites are right in the Panel’s Settings, or on the mobile page. Scan the QR, and you’re paired. Star the repo or file an idea at artokun/comfyui-mcp.