> ## 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.
# Local vs. Comfy Cloud agent
> Comfy-Org's In-App Agent and Comfy Cloud MCP run on Comfy Cloud; comfyui-mcp runs on your own ComfyUI install, on any LLM — Claude or ChatGPT on your existing subscription, or a free local model via Ollama with no account and no network. An honest comparison of when to use which, including the cases where Comfy Cloud is the better answer.
If you're deciding between Comfy-Org's official agent tooling and this project,
the short version is that **they solve different problems** — and for a decent
number of people the official option is the right call. This page lays out both
honestly so you can pick once and move on.
Comfy-Org ships fast, and the statuses below are as of **July 2026**. Check
[docs.comfy.org/agent-tools](https://docs.comfy.org/agent-tools) for the current
state before treating any of this as settled.
## The one-line answer
**No local GPU, or you want zero setup?** Use Comfy Cloud. It's first-party,
it's maintained by the team that builds ComfyUI, and you don't have to think
about drivers, VRAM, or model downloads.
**Already running ComfyUI on your own machine?** That's what this project is
for. Your GPU is already paid for, your models are already on disk, and your
Claude or ChatGPT subscription already covers the reasoning.
## What Comfy-Org offers
Comfy-Org's [Agent Tools](https://docs.comfy.org/agent-tools) page lists three
relevant products:
| Product | What it is | Status (Jul 2026) |
| ---------------------- | --------------------------------------------------------------------- | ------------------------------------------- |
| **Comfy Cloud MCP** | Hosted MCP server; generation runs on Comfy Cloud GPUs | in public beta |
| **Comfy In-App Agent** | Chat-driven agent inside Comfy Cloud that builds and edits your graph | in private alpha |
| **Comfy Local MCP** | First-party local MCP server driving your own ComfyUI install | in private test; not publicly available yet |
Comfy Cloud MCP is genuinely good at what it does: it exposes the ComfyUI engine
over MCP, signs in with OAuth from Claude Code, Claude Desktop, and Codex, and
gives you managed access to models like Kling, Veo, Seedance, Flux, and
ElevenLabs without owning a GPU.
## What this project offers
`comfyui-mcp` is an MCP server plus a sidebar agent that drives **your** ComfyUI
install — local, LAN, VPS, or RunPod. 38 MCP tools, 42 model-specific skills,
and a panel agent that edits your live graph in natural language.
The structural difference is **where the reasoning comes from**. This project is
provider-agnostic: the same tools and the same panel run on Claude, ChatGPT,
Gemini, Grok, or Kimi using a subscription you already pay for, on any
OpenAI-compatible endpoint, or on a free local model through Ollama, LM Studio,
or llama.cpp — with no account and no outbound network at all.
## Side by side
| | Comfy Cloud (In-App Agent / Cloud MCP) | comfyui-mcp |
| --------------------------------------- | ------------------------------------------ | ---------------------------------------- |
| Generation runs on | Comfy Cloud GPUs | Your GPU (or your VPS / RunPod pod) |
| Requires a Comfy account | Yes | No |
| Works fully offline | No | Yes, with a local model |
| Reasoning model | Provided by the service | **Yours** — any LLM, incl. free local |
| Cost model | Cloud credits | Your existing subscription, or \$0 local |
| Your images/prompts leave the machine | Yes | Only if you choose a hosted LLM |
| Manages your custom nodes & model files | N/A (managed environment) | Yes — installs nodes, downloads models |
| Availability | Cloud MCP public beta; agent private alpha | Available now |
| First-party support | Yes | Community project |
## Choose Comfy Cloud if…
* You don't have a GPU, or yours is too small for the models you want
* You want it to work with no setup and no maintenance
* You'd rather buy credits than manage CUDA, VRAM, and checkpoint downloads
* Your team is already on Comfy Cloud and you want one vendor and one invoice
* You want first-party support from the people who build ComfyUI
These are good reasons. If several apply to you, stop reading and go use
[Comfy Cloud](https://comfy.org/) — you'll have a better time.
## Choose comfyui-mcp if…
* **You already have a GPU.** You've paid for it; cloud credits are a second bill
for compute you already own.
* **Your work can't leave the machine.** Client material under NDA, licensed
training data, anything you can't upload. With a local model, nothing does.
* **You want to use the LLM subscription you already have.** Your Claude or
ChatGPT plan already covers the reasoning — no second meter for the agent.
* **You want it free, or you want it offline.** Ollama, LM Studio, and llama.cpp
run with no account and no network. See [Local LLMs](./local-llms).
* **Your install is the point.** Your custom nodes, your LoRAs, your checkpoint
collection. The agent [installs nodes and downloads models](./tools/models) into
the install you actually use.
* **You want to choose the model.** Different LLMs are better at different
things; the [LLM Arena](./arena) scores them on real ComfyUI tasks so you can
pick on evidence. See [Backends](./backends).
## "Won't the official local MCP make this redundant?"
Fair question, and worth answering directly rather than pretending otherwise.
Comfy Local MCP is a **first-party driver for a local install**, and when it
ships it will be a good one. But it is a different layer than this project: it
connects an agent to ComfyUI, where `comfyui-mcp` also brings the agent, the
provider-neutral backend that runs it across eleven providers and local
runtimes, the sidebar UI, the model-specific skills, and the node/model
management.
The honest read: if all you want is "let Claude Code talk to my local ComfyUI,"
first-party tooling will eventually cover that, and you should probably use it.
If you want to bring your own model — including a free local one — and have the
agent live inside ComfyUI and manage the install itself, that's the gap this
project exists to fill.
They also compose. This project can target Comfy Cloud too (set
`COMFYUI_API_KEY` — see [Cloud deployment](./cloud-deployment)), so one config
covers a local install, a LAN box, a VPS, and Comfy Cloud.
## Getting started
MCP server running against your local ComfyUI in about a minute.
Fully offline on Ollama, LM Studio, or llama.cpp — no account.
Claude, ChatGPT, Gemini, Grok, Kimi, or any OpenAI-compatible endpoint.
The sidebar agent that drives your live graph.