> ## 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. # Roadmap > What's shipped, what's being built, what's next — and what we've decided not to build, with the reasoning. Ordered by priority, not by date. ## What this covers **The whole project, not one repository.** comfyui-mcp is several pieces that ship separately and version independently: | Piece | Repo | What it is | | ----------------------- | --------------------------------------------------------------------- | ----------------------------------------------------- | | **MCP server + plugin** | [`comfyui-mcp`](https://github.com/artokun/comfyui-mcp) | the tools and the orchestrator — the control plane | | **Panel** | [`comfyui-mcp-panel`](https://github.com/artokun/comfyui-mcp-panel) | the agent sidebar inside ComfyUI | | **Mobile app** | [`comfyui-mcp-mobile`](https://github.com/artokun/comfyui-mcp-mobile) | chat, queue, LoRAs and desktop-tab mirroring | | **Relay** | [`comfyui-mcp-relay`](https://github.com/artokun/comfyui-mcp-relay) | Cloudflare Worker for a stable `wss://` remote bridge | Items below are labelled by piece where it matters. Everything is planned as one product even though it lands in four places. The version numbers in [How it got here](#how-it-got-here) are the **MCP server's**. The panel and the mobile app carry their own, unrelated version lines — panel 0.9.x is not "behind" server v0.44. Don't compare them. ## How to read this **There are no dates on this page, and that's deliberate.** This project ships fast and reorders often — the RunPod arc jumped the queue in a single week. A roadmap with quarters on it would be wrong within days, and a wrong date is worse than no date: it turns a plan into a promise nobody made. So this page commits to **order, not timing**. "Next" means next in line, not next month. If something here matters to you, say so in [Discord](https://discord.gg/cW9arBhzCu) — what people actually do day to day moves things up this list far more reliably than the size of the request. The most useful section is probably the last one. **[Not planned](#not-planned)** says what we've decided *against* and why, so you can stop waiting for it. *** ## Shipped Everything here works today and is covered by the docs. | Area | What you get | | ------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Agent in ComfyUI** | The panel puts an agent beside your canvas — it reads the graph, edits nodes, runs workflows, and explains failures. See [Panel](./panel). | | **Bring your own model** | Claude, ChatGPT/Codex, Gemini, Antigravity, Grok, Kimi, GLM, Ollama, LM Studio, llama.cpp, OpenRouter, or a custom endpoint — one picker, one orchestrator. See [Backends](./backends). | | **Local models** | Fully local operation through Ollama / LM Studio / llama.cpp, including small models. See [Local LLMs](./local-llms). | | **Workflow authoring** | Build, patch, slice, validate and lock workflows from natural language; import from an image's metadata. | | **Models & custom nodes** | Search, install and repair node packs; find and download models — including working out which models a workflow is missing and which quantisation fits your VRAM. | | **Civitai browser** | Browse images, videos and models in-panel, filter by base model and creator, and pull a workflow straight onto the canvas. | | **Mobile app** | Chat to the agent, watch the queue, browse LoRAs, and mirror a desktop tab to drive your rig from your phone. See [Mobile](./mobile). | | **Apps (micro-apps)** | Turn a workflow into a one-click app with a real form: convert and publish in the panel, run it locally or on a pod, install others' apps from the public registry — and run the same apps from the phone's Apps tab or an agent via the `apps_*` tools. See [Apps](./apps). | | **Remote & cloud** | Point the panel at a remote ComfyUI, a self-hosted relay, or Comfy Cloud. See [Remote connector](./remote-connector). | | **RunPod pods** | Create, connect to and stop RunPod pods from the panel, with billing-safe defaults and an honest host indicator. | | **LoRA training (local)** | Character LoRAs on FLUX.1-dev, driven by the agent through `train_*` tools in a GPU Docker container — dataset in, `.safetensors` in `models/loras/` out. | ## Building now Actively in progress. These are the things most likely to move next. | Item | Where it stands | | ----------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------- | | **LoRA training on RunPod pods** | The transport landed; follow-up work continues on idle-stop scoping and a dockerless native path for machines without Docker. | | **Training beyond character LoRAs** | Style, slider and edit flows, and base models past FLUX.1-dev. The config layer was built to broaden; the flows aren't written yet. | ## Next Queued and specified, not started. Ordered. | Item | Why it's here | | ------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **Honest previews in the mobile app** | The **Training** tab is still inert placeholder UI — six cards that read as a shipped feature and have already misled people into asking how far along it is. Either label it plainly or hide it — the cheapest fix on this list. (The **Apps** tab was the other half of this item and is now real; see [Apps](./apps).) | | **Mobile training** | Wire the mobile Training tab to the trainer that already works. The hard part is done; this is mostly plumbing, and it retires the item above. | | **Tablet layout** | Real, unglamorous responsive work. Worth doing once the screens above have something worth showing. | ## Under consideration Wanted, but blocked on a design question rather than on effort. No commitment. ### Multi-user sessions Letting someone else connect to *your* session and drive *your* machine — pair debugging a broken workflow, or a shared studio rig. Worth being precise about the blocker, because it isn't the obvious one: **the transport already exists.** Pairing, socket fan-out and per-session auth were all built for phone-to-desktop mirroring, which is the same plumbing. What's missing is a **trust model**. Who is allowed to drive? What can they see of your canvas, your history, your file system? How do you revoke access from someone mid-run? Those questions need answering before any of it gets built — shipping a half-designed answer on someone else's GPU is how people get hurt. *** ## Not planned Decided against, with the reasoning, so nobody waits on them. Not permanent — a good argument reopens any of these — but nothing here is queued. ### Dates and version-numbered milestones See [the top of this page](#how-to-read-this). Order, not timing. ### Voice-only control (smart glasses, headsets) Driving the desktop agent by voice from glasses comes up regularly, and the plumbing would mostly reuse the mirroring transport. The reason it isn't queued is that a voice-only surface **has no canvas**. There's no preview to glance at, no way to resolve *"use the other LoRA"* when two are loaded, and no visible undo when the agent does the wrong thing to your graph. Those are the actual problems, and adding a microphone doesn't solve any of them. Until there's an answer, this would ship as a demo that's unpleasant to rely on. ### Smartwatch Everything above applies, on a smaller screen with a tighter power budget. Nobody has yet described a job it does better than the phone already does. ### A hosted service that runs ComfyUI for you This project is a **local-first control plane**. It connects to your ComfyUI — on your machine, your remote box, your pod, or Comfy Cloud — and it doesn't want your files or your prompts. Running inference on your behalf is a different product with a different trust model, and it isn't this one. *** ## How it got here The arc so far, for context on where the next items sit. This is deliberately coarse — for what changed in a specific release, see the [Changelog](./changelog). Public releases begin at **v0.2.0**. There is no v0.0.1 or v0.1.x — the first two weeks of work landed straight on `main` before the package was published, starting from the initial commit on **2026-02-15**. Releases shown are the **MCP server's** — the panel and mobile app version separately, and the eras below describe the project as a whole. | Era | Server releases | What it was about | | ----------------------------------------- | --------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | **The MCP server**
Feb 2026 | v0.2 – v0.3 | The original idea: drive ComfyUI from an agent. Image generation, workflow analysis, a UI→API workflow converter, hierarchical visualisation, model skills, and the shift from blocking runs to a non-blocking queue with background progress and process control. | | **Quiet stretch**
Mar – Apr 2026 | v0.3.x | Three feature commits in two months. Remote ComfyUI over HTTP, subgraph/component expansion in the converter, HTTPS/WSS support. Groundwork, not headlines. | | **Environment & lifecycle**
May 2026 | v0.4 – v0.8 | Making it survive real installs: custom-node verify/scaffold with CI, installs pinned to a git ref, auth for gated and private models, declarative environment manifests, cloud storage providers, server-side image conversion, and cached node-skill discovery. The embedded panel was designed at the end of this stretch. | | **The panel era**
Jun 2026 | v0.9 – v0.22 | The biggest month by volume (\~102 feature commits). The agent moved *into* ComfyUI beside the canvas, gained multiple providers, and grew the infrastructure to run anywhere: one-click installer packs, a RunPod image, a secure `wss://` bridge with an opt-in relay, and live reasoning visibility. | | **Reach and depth**
Jul 2026 | v0.23 – v0.44 | Two directions at once. **Reach:** the mobile app, desktop-tab mirroring, the Civitai browser, more providers (Antigravity, GLM, Kimi), and compact tool mode so small local models can drive the tools. **Depth:** the LLM Arena, first-party RunPod pods with billing-safe defaults, LoRA training, and missing-model resolution that accounts for your VRAM. | Two things that table won't tell you. The pace is uneven on purpose — a two-month lull in spring, then \~180 feature commits across June and July. And the ordering has never been fixed in advance: RunPod and training both jumped the queue because they turned out to matter more than what was nominally next. That's the main reason this page carries no dates. ## Something missing? If you're waiting on something that isn't listed, or one of the "not planned" calls looks wrong for how you actually work, open an issue on [GitHub](https://github.com/artokun/comfyui-mcp/issues) or say so in [Discord](https://discord.gg/cW9arBhzCu). Requests that describe **what you're trying to do** are far more useful than ones that name a feature — several items on this page are ordered the way they are because someone explained their day, not because they asked loudest.