ComfyUI Download and Install (2026): Desktop vs Portable, the Built-In Manager, Your First Image

ComfyUI is the free, open-source app that runs AI image, video and audio models on your own computer, by connecting boxes called nodes into a workflow. The quickest ComfyUI download for most people is Comfy Desktop for Windows, macOS on Apple Silicon or Linux; Windows users with an NVIDIA, AMD or Intel card can also use the Portable build, and everyone else can install it by hand with Python. The current release is v0.39.0, from October 5, 2026. Here is the part most install guides still get wrong: ComfyUI Manager is now built into ComfyUI itself. You no longer clone it from GitHub. On Desktop it is already switched on, and on Portable and manual installs it is switched off until you install its requirements and start ComfyUI with one flag, --enable-manager.

Jake runs a phone repair shop, and after a month of making window posters in the Gemini app he wanted something he could use as much as he liked, without daily limits and without uploading photos of customers' phones anywhere. His nephew had left an old gaming PC in the back room, and Ethan, who sets up computers and cloud accounts for a few local businesses, offered to get ComfyUI running on it one evening. Two things went wrong in the first ten minutes: a guide from last year told them to download ComfyUI Manager separately, and the first workflow opened with a red warning about a missing model. Neither was a real problem. This page is that evening written down: which ComfyUI download to pick, how to install it on Windows, Mac and Linux, including Kali, how to turn on ComfyUI Manager, where models go, how to make your first image, how to update safely, and how to fix the errors that stop people on day one.

If ComfyUI looks intimidating in screenshots, all those wires and boxes, it helps to know you do not have to build any of it to start. ComfyUI ships with ready-made workflow templates for the popular models, and your first image comes from one of them. The boxes start to make sense after a few runs, the way a recipe does once you have cooked it twice.

What is ComfyUI, and is it free?

ComfyUI is a node-based interface and engine for generative AI. Each node does one job: load a model, turn your prompt into something the model understands, generate, decode, save. You connect them with wires, and the chain is called a workflow. Because every step is visible, you can see exactly what happens and swap any part of it, which is why ComfyUI became the place new open models arrive first.

It runs image models such as Stable Diffusion 1.5, SDXL, SD3.5, Flux.1, Flux.2, Qwen Image, Z-Image, Hunyuan Image 2.1, HiDream, Lumina Image 2.0 and Chroma, as well as video and audio models, including ACE-Step 1.5, Stable Audio 3 and YuE 2 for music. The project is developed by Comfy Org on GitHub, where it has more than 136,000 stars, and it is licensed under GPL-3.0.

Is ComfyUI free? The software is free, with no account needed, and everything you generate with local models costs nothing beyond electricity. Two optional things cost money:

• Comfy Cloud runs ComfyUI in a browser on Comfy's own GPUs, for people without a capable computer. Signing in with Google gives 400 free credits a month; heavier use is a paid plan.
• Partner Nodes call closed, hosted models from inside a workflow, such as OpenAI, Kling, Luma or Google's Nano Banana 2.1, which has just been added. They need a Comfy account and prepaid credits, and they send your inputs to that outside service. If you want a purely local setup, two newer launch flags help: --disable-partner-nodes hides them, and --offline keeps ComfyUI from reaching the network. The older --disable-api-nodes flag is deprecated.

ComfyUI follows a weekly release cycle that usually targets Monday, and roughly every two weeks a new stable version is tagged. Comfy Desktop builds from the latest stable core, and the frontend, the part you see in the browser, is developed in its own repository and merged into ComfyUI every couple of weeks.

Which ComfyUI download should you choose?

There are four ways to install ComfyUI locally, plus the cloud. The right one depends on your system and how much you like the command line:

Two honest notes about Portable. ComfyUI's own README says the portable build is "not recommended for regular users" and points them to the desktop app instead. And Portable has four separate downloads: one for modern NVIDIA cards (the 20 series and newer), one for older NVIDIA cards (the 10 series, with Python 3.12 and CUDA 12.6), one for AMD and one for Intel. Picking the wrong one is the most common reason Portable will not start.

The official places to download are comfy.org/download for Desktop and the ComfyUI GitHub releases page for Portable. Anything else is a mirror at best.

ComfyUI system requirements

ComfyUI itself is small. The models are what need space and memory. The current requirements:

• Operating system: Windows, Linux, or macOS on Apple Silicon (M1 to M4).
• Disk: at least 4.85 GB per Comfy Desktop installation, before models. A single modern image model is often several gigabytes, and video models far more, so plan for tens of gigabytes.
• Memory: 8 GB of RAM minimum for Comfy Desktop, 16 GB recommended.
• GPU: an NVIDIA or AMD card is recommended but not required. ComfyUI supports NVIDIA (CUDA 13.0 PyTorch), AMD on Linux (ROCm 7.2), AMD on Windows (ROCm 10.0 through AMD's multi-architecture PyTorch packages, covering RDNA 2, 3, 3.5 and 4), Intel Arc (PyTorch with torch.xpu), Apple Silicon (Metal), plus Ascend, Cambricon and Iluvatar accelerators. CPU-only works with --cpu, very slowly.
• Python (manual installs): 3.13 is recommended. 3.14 works, but some custom nodes may have issues. 3.12 is the fallback if a custom node's dependencies fail on 3.13.
• PyTorch: 2.7 or newer is supported, and newer is strongly recommended. NVIDIA 20-series and newer cards need a cu130 or newer PyTorch build.
• Browser: Google Chrome 143 or later gives the best experience; versions 142 and below have known visual glitches with ComfyUI.

How much graphics memory you need depends entirely on the model. Small Stable Diffusion 1.5 models run on modest cards; the newest image and video models want far more. Our laptop tiers guide explains what each memory size can realistically run, and the LTX-2.5 guide shows how quickly video models raise the bar.

Jake's back-room PC turned out to have a GeForce GTX 10-series card. That one detail decided the download: the older-NVIDIA Portable build, or the Desktop app.

How to install Comfy Desktop on Windows, Mac and Linux

Comfy Desktop is a launcher that installs, manages and starts ComfyUI for you, including its own Python environment. You do not install ComfyUI separately; each installation you create inside the app runs on its own, with its own version, custom nodes and settings. That makes it easy to keep one stable setup for work and a second one for trying new things.

On Windows:

1. Go to comfy.org/download; the page detects your platform.
2. Run the downloaded .exe installer and follow the prompts.
3. Launch Comfy Desktop from the Start menu. On the Welcome screen, create your first installation.
4. Wait while it sets up Python and the dependencies, then open the installation. ComfyUI Manager is already included and switched on.

On macOS: Comfy Desktop needs macOS 13 or later on Apple Silicon. Install it like any other Mac app. If macOS says the app "is damaged", allow it in System Settings under Privacy and Security; if it runs slowly, granting it Full Disk Access in Privacy settings helps.

On Linux: download the official .deb or AppImage from the same link (x64, Ubuntu 22.04 or later recommended), then install it:

# Debian or Ubuntu package
sudo apt install ./*.deb

# or the AppImage
chmod +x ./*.AppImage
./*.AppImage

If the app will not open on Ubuntu 24.04 or newer, the distribution is probably blocking Electron's sandbox through AppArmor. Starting it with --no-sandbox works around that, at the cost of turning off a security layer, so use it knowingly.

Desktop updates itself: a Desktop Update Ready button appears when a new version is out, and you can toggle automatic updates under Desktop Settings, then Updates. Updating the app is separate from updating the ComfyUI engine inside each installation, which you do from that installation's Manage panel, choosing the Stable channel (the default) or Latest on GitHub.

If you used the older single-install desktop app, now called Legacy Desktop, Comfy Desktop detects it and migrates your custom nodes, workflows, models and settings automatically, leaving the old install untouched as a backup. Models are linked rather than copied, so do not delete the old model folder afterwards without checking.

Uninstalling the app from Settings, Apps, Installed apps removes only the launcher. Your installations, models and outputs stay on disk until you delete them:

Those folders hold your models and your pictures, so treat them as personal data, not leftover program files.

How to install ComfyUI Portable on Windows

ComfyUI Portable is the complete program in one folder, with its own embedded Python, so nothing is installed into Windows. It always runs the latest code, which is why tinkerers like it.

1. Download the right build from the ComfyUI GitHub releases page: ComfyUI_windows_portable_nvidia.7z for NVIDIA 20 series and newer, the _nvidia_cu126 build for 10-series cards, the _amd build or the _intel build.
2. Extract it with 7-Zip, or with File Explorer on recent Windows versions. If extraction fails, right-click the file, choose Properties and tick Unblock.
3. Update your graphics driver first. The NVIDIA build ships with Python 3.13 and PyTorch for CUDA 13.0, and an old driver is the usual reason it does not start.
4. Double-click the matching launcher: run_nvidia_gpu.bat, run_amd_gpu.bat or run_intel_gpu.bat, or run_cpu.bat for CPU only.
5. Wait for the line To see the GUI go to: http://127.0.0.1:8188. Your browser normally opens by itself; if not, open that address.
6. Leave the black command window open. Closing it stops ComfyUI.

The extracted folder looks like this, and knowing it saves a lot of searching later:

ComfyUI_windows_portable
├── ComfyUI              the program (models, custom_nodes, output live inside)
├── python_embeded       the private Python environment
├── update               update scripts
├── README_VERY_IMPORTANT.txt
├── run_cpu.bat
└── run_nvidia_gpu.bat   (or run_amd_gpu.bat / run_intel_gpu.bat)

Jake and Ethan went with the 10-series Portable build on the old PC. It started on the second try, after a driver update.

How to install ComfyUI manually (Windows, Linux, macOS)

The manual install is the most flexible and the only route for some hardware. It has four steps: make an isolated Python environment, get the code, install PyTorch for your GPU, install ComfyUI's requirements.

1. Create a virtual environment. ComfyUI's dependencies can clash with other Python software, so keep it separate. With Python's built-in venv:

python -m venv comfy-venv
# Windows
comfy-venv\Scripts\activate
# Linux and macOS
source comfy-venv/bin/activate

Conda works too (conda create -n comfyenv, then conda activate comfyenv). On Windows, install the Microsoft Visual C++ Redistributable if you do not have it.

2. Get the code with Git:

git clone https://github.com/Comfy-Org/ComfyUI.git
cd ComfyUI

3. Install PyTorch for your hardware. This is the step that decides whether your GPU is used at all:

# NVIDIA
pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu130

# AMD on Linux (ROCm 7.2)
pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm7.2

# Intel Arc (Windows and Linux)
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

AMD on Windows uses AMD's own ROCm 10.0 packages, on Windows 11 with a current AMD driver and 64-bit Python 3.13. The install command uses a device-all extra that covers every supported card; you can swap it for your card's target, such as device-gfx1100 for an RX 7900 XT or XTX, to download less:

pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "torch[device-all]==2.13.0+rocm10.0.0" "torchvision[device-all]==0.28.0+rocm10.0.0" "torchaudio==2.11.0.2+rocm10.0.0"

On an Apple Silicon Mac, install the latest PyTorch nightly following Apple's Metal PyTorch guide, then check that the GPU backend is available. ComfyUI uses PyTorch's MPS backend on Macs, not MLX:

python -c "import torch; print(torch.backends.mps.is_available())"

4. Install ComfyUI's requirements and start it:

pip install -r requirements.txt
python main.py

Then open http://127.0.0.1:8188. python main.py --help lists every startup option.

Prefer one command? comfy-cli does the whole thing:

python -m venv venv
source venv/bin/activate        # Windows: venv\Scripts\activate
pip install comfy-cli
comfy install

How to install ComfyUI on Kali Linux

Kali is Debian-based, so both Linux routes apply: the official .deb with sudo apt install ./*.deb, or the manual install above. Three Kali habits save time:

• Always use a virtual environment. Kali, like current Debian, refuses system-wide pip install with an "externally-managed-environment" error. The venv in step 1 is the right fix, not --break-system-packages. Our Kali pip guide explains why.
• Check the GPU before ComfyUI. If nvidia-smi does not show your card, ComfyUI will not either. Confirm PyTorch sees it with python -c "import torch; print(torch.cuda.is_available())".
• A virtual machine is not a GPU machine. Kali in VirtualBox or VMware normally cannot use your graphics card for CUDA, so ComfyUI falls back to the CPU. For real use, install on bare metal or a system with direct GPU access.

If you hit missing-library errors such as libcuda.so.1: cannot open shared object file, see the troubleshooting section below; it is a library path issue, and the fix is the same on Kali as on Ubuntu.

ComfyUI Manager: how to install, enable and update it

ComfyUI Manager is the extension that installs, updates and removes custom nodes, and helps fetch models, from inside ComfyUI. It used to be a separate project you cloned into the custom_nodes folder. The new ComfyUI Manager is built into ComfyUI core and installed as its own Python package, pinned in a file called manager_requirements.txt. How you turn it on depends on how you installed ComfyUI.

Comfy Desktop: nothing to do. Every installation includes Manager and starts with --enable-manager. Per installation, under the Manage panel's Startup Args tab, you choose the Manager Security Level (Strict, Standard, which is recommended, Relaxed or Permissive) and the Manager Network Mode (Public by default, Private, Offline or Personal cloud).

ComfyUI Portable: open a terminal in the ComfyUI_windows_portable folder and run:

.\python_embeded\python.exe -m pip install -r ComfyUI\manager_requirements.txt
.\python_embeded\python.exe -s ComfyUI\main.py --windows-standalone-build --enable-manager

To make it permanent, copy run_nvidia_gpu.bat to a new file and add --enable-manager to the end of its launch line, then start ComfyUI with that file.

Manual install: with your virtual environment active:

pip install -r manager_requirements.txt
python main.py --enable-manager

Updating ComfyUI Manager catches people out: updating ComfyUI can leave the installed Manager package unchanged. Run the manager_requirements.txt install again in the same environment you launch ComfyUI from, then restart. Desktop keeps Manager current with its own updates.

The old way still exists. If you follow an older guide and clone Manager into ComfyUI/custom_nodes/comfyui-manager, it works, but the folder must be exactly that path. Do not unzip it straight into custom_nodes, and do not end up with ComfyUI-Manager/ComfyUI-Manager or ComfyUI-Manager-main; those nested names are the classic reason the Manager button never appears.

Once Manager is on, installing a custom node is: open Manager, search for the node pack, install, restart ComfyUI. Custom nodes are other people's Python code running on your computer, so the Standard security level and the official registry are there for a reason. Ethan's rule for Jake was simple: install a node pack only when a workflow you trust asks for it.

How to install ComfyUI custom nodes safely

Custom nodes are community extensions that add new abilities to ComfyUI: support for a new model, better upscaling, video tools, face fixes and hundreds more. Every custom node install has the same two parts: the node's code goes into ComfyUI/custom_nodes, and its Python dependencies go into ComfyUI's own environment. There are three ways to do it:

For a manual install, clone the node into the folder, then install its requirements with the Python that runs ComfyUI. On Portable, that is the embedded Python, not the one installed in Windows:

cd ComfyUI/custom_nodes
git clone https://github.com/example/some-node-pack

# Portable (run from the ComfyUI_windows_portable folder)
python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\some-node-pack\requirements.txt

# Manual install (virtual environment active)
pip install -r some-node-pack/requirements.txt

Installing dependencies into the wrong Python is the single most common reason a freshly installed node shows up red. Restart ComfyUI after any install.

ComfyUI's own documentation is blunt about safety: custom nodes are code, and a malicious one can compromise your system. Install node packs from trusted authors that the community actually uses, read the README before installing, and avoid obscure packs from unknown sources. If ComfyUI starts misbehaving after a new node, --disable-all-custom-nodes tells you in one restart whether a node is to blame.

Your first image in ComfyUI, step by step

With ComfyUI open in the browser, the fastest route to a picture:

1. Open a template. Use the Templates browser, or load the default text-to-image workflow with Ctrl + D.
2. Let ComfyUI fetch the model. Most installs ship without base models. When the workflow needs one you do not have, ComfyUI shows a missing-model warning with a download link. For the basic workflow that is v1-5-pruned-emaonly-fp16.safetensors, a Stable Diffusion 1.5 checkpoint, which goes in ComfyUI/models/checkpoints.
3. Write your prompt in the positive prompt box, and what you do not want in the negative box.
4. Run it with the Run button or Ctrl + Enter. The result appears in the Save Image node and is written to ComfyUI/output.
5. Change one thing at a time, such as the prompt, the seed or the image size, and run again. Ctrl + Shift + Enter puts a run at the front of the queue, and Ctrl + Alt + Enter cancels the current one.

That red warning Jake and Ethan saw on the first workflow was step 2. Clicking it showed the model and the link; four minutes of downloading later, the first picture of a repaired phone on a workbench was on screen.

Where models go. Each model type has its own folder under ComfyUI/models/: checkpoints for full models, loras, vae, controlnet, embeddings and upscale_models, among others. Many newer models come as several files that each go in a different folder, so follow the model's own instructions. If you copy files in while ComfyUI is open, restart or refresh so the lists update. To remove a model, delete its file; there is no uninstall button for models yet.

Sharing models with another app, such as an older Stable Diffusion web UI, or between several ComfyUI installs, is done with a config file instead of copies. Rename extra_model_paths.yaml.example in the ComfyUI folder to extra_model_paths.yaml, point base_path at the other folder, and restart:

my_models:
    base_path: D:/AI/models/
    checkpoints: checkpoints
    loras: loras
    vae: vae
    controlnet: controlnet

On Comfy Desktop, add model folders under the Storage tab of the capsule menu at the top of the app instead.

ComfyUI workflows: templates, loading from a picture, and models to try

A workflow is the whole graph of nodes, saved as a JSON file. You can save and open them with Ctrl + S and Ctrl + O, or from the Workflows menu. Two features make sharing them easy. First, ComfyUI can recover a complete workflow, including the seed, from the images and media it generated, so a picture someone shares can carry its own recipe; drop it onto the canvas to load it. Second, the built-in template library, also browsable at comfy.org/workflows, holds maintained, ready-to-run workflows for the models ComfyUI supports.

That support is broad. ComfyUI runs, among others:

• Image generation: Stable Diffusion 1.5, SDXL, SD3.5, Flux.1, Flux.2, Qwen Image, Z-Image, Hunyuan Image 2.1, HiDream, Lumina Image 2.0 and Chroma.
• Image editing: Flux Kontext, Flux.2 Klein, Qwen Image Edit, HiDream E1.1, OmniGen2 and Boogu.
• Video: Wan 2.1 and 2.2, LTX-Video 2 and 2.3, HunyuanVideo 1.5, CogVideoX and Mochi.
• Audio: ACE-Step 1.5, Stable Audio 3, MiniMax Music 3 and YuE 2.
• 3D and vision: Hunyuan3D 2.1, SAM 3 and Depth Anything 3, plus upscalers such as SeedVR2 and SUPIR.
• Text: Gemma 3 and 4, Qwen3, Qwen3.5 and Qwen3-VL, including image inputs.

Where to start depends on your graphics card. On a modest card, the Stable Diffusion 1.5 workflow from the first-image section is the right teacher. With more memory, Qwen-Image-2.1 handles both generation and editing in one model, and our Qwen-Image-2.1 guide walks through it in ComfyUI. Z-Image-Turbo, a distilled 6-billion-parameter model built for speed, is a good next step for fast results, and video models such as Wan 2.2 and LTX-2.3 are where large cards earn their keep.

One reassurance for privacy-minded users: ComfyUI's core runs fully offline and does not download anything unless you ask it to. Model downloads, Manager installs and Partner Nodes are the only things that reach out, and --offline turns off the paid nodes and keeps built-in features local.

ComfyUI keyboard shortcuts worth learning first

On a Mac, use Cmd in place of Ctrl.

How to update ComfyUI without breaking it

An update has two parts: ComfyUI's own code, and its core dependencies, which include the frontend package, the workflow templates and the node help pages. Most "the update broke it" stories come from updating only the first part.

Portable: use the scripts in the update folder.

• update_comfyui.bat updates to the latest development code and its dependencies. Use this one day to day.
• update_comfyui_stable.bat updates to the latest stable release.
• update_comfyui_and_python_dependencies.bat also reinstalls PyTorch and every Python dependency. ComfyUI's own docs label it dangerous: it can break custom nodes that need specific package versions. Use it only to fix dependency problems, and back up the whole folder first.

Desktop: the app updates itself; update the ComfyUI engine in each installation from its Manage panel.

Manual: activate the virtual environment, then both commands, always together:

cd ComfyUI
git pull
pip install -r requirements.txt

If you only run git pull, the frontend and templates stay old, and you will see missing menus, missing templates, or a startup line saying Falling back to the default frontend. Running the requirements install in the right environment fixes it. To go back to an earlier version, git log --oneline shows the history, git checkout with a commit ID moves to it, and git checkout master returns to the latest.

Ethan's habit, which he passed on to Jake: before any update, copy the custom_nodes folder and your saved workflows somewhere safe. It takes a minute and turns a bad update into a short inconvenience.

ComfyUI not working? The common errors and their fixes

ComfyUI's own team says most reported problems come from custom nodes, so the first test for almost anything is to start ComfyUI with --disable-all-custom-nodes. If the problem disappears, a custom node is the cause; turn them back on in halves until you find it. For everything else:

1. "Torch not compiled with CUDA enabled." PyTorch was installed without GPU support. Run pip uninstall torch, then reinstall with the cu130 command from the manual install section, and check with python -c "import torch; print(torch.cuda.is_available())".
2. Portable will not start, or closes at once. Update the graphics driver, and check you downloaded the build for your card; 10-series NVIDIA cards need the cu126 build.
3. "Prompt execution failed". Click Show report to read the real error. It usually names a missing model, a node from an uninstalled pack, or running out of graphics memory.
4. Out of memory or very slow. Lower the resolution or batch size, close other GPU programs, and try --lowvram (text encoders on the CPU) or --novram if that is not enough. --reserve-vram 2 keeps 2 GB free for the system, and --preview-method none saves a little more.
5. The address will not load, or the port is taken. The default is port 8188. Start with --port 8189 to use another.
6. Models do not appear in the list. Check the file is in the right subfolder under ComfyUI/models/, then restart or refresh. For shared folders, check the paths in extra_model_paths.yaml.
7. Linux: libcuda.so.1, libnccl.so or libnvinfer.so "cannot open shared object file". PyTorch cannot find NVIDIA libraries. Add the nvidia/.../lib folders inside your environment's site-packages to LD_LIBRARY_PATH, or check the system CUDA path with ldconfig -p | grep cuda.
8. Older AMD cards on Linux. For RDNA2 cards such as the 6700 or 6600, try HSA_OVERRIDE_GFX_VERSION=10.3.0 python main.py; for RDNA3 cards such as the 7600, HSA_OVERRIDE_GFX_VERSION=11.0.0.
9. Linux Desktop app will not open. Ubuntu 24.04 and newer may block Electron's sandbox; start the AppImage with --no-sandbox.
10. macOS says the app is damaged. Allow it in System Settings, Privacy and Security.
11. The Manager button is missing. On Portable or manual installs, you have not started with --enable-manager, or its requirements are not installed. With an old cloned Manager, check the folder is exactly custom_nodes/comfyui-manager.
12. Partner Nodes fail. They need you to be signed in to a Comfy account with more than zero credits, and login only works from 127.0.0.1 or localhost, not from another computer on your network.
13. Visual glitches in the interface. Update Chrome to version 143 or later.

If none of these fit, ComfyUI's GitHub issues are the place to report it. Include your install type, ComfyUI version, GPU, driver version and the full report text; a screenshot of a red box alone is rarely enough to help.

Comfy MCP: let an AI agent drive ComfyUI

A newer option, for people who already use AI coding assistants: Comfy MCP connects AI agents to ComfyUI over the Model Context Protocol. Once connected, you can ask an agent in plain language to generate an image, find a template, inspect nodes or run a saved workflow, and it calls ComfyUI for you. There are two connections.

• The cloud connection links the agent to a Comfy Cloud account, with nothing to install; it lives at https://cloud.comfy.org/mcp. In Claude Code, for example, it is added with claude mcp add --transport http comfy-cloud https://cloud.comfy.org/mcp, followed by a browser sign-in. Comfy recommends it for new users and for Mac users who want large open models.
• The local connection, comfy-mcp, is Comfy's first-party, open-source server for driving your own ComfyUI install. It needs comfy-cli version 1.14.0 or newer on your PATH and a running ComfyUI (start it with comfy launch):

pip install "comfy-cli>=1.14.0"
pip install comfy-mcp

# example: register it with Claude Code
claude mcp add comfy-mcp -e COMFY_BIN=/path/to/venv/bin/comfy -- comfy-mcp

The local server can check whether ComfyUI is running, start and stop it, list and inspect the nodes in your live install, custom nodes included, and run workflows. The two connections sign in separately, and running both is normal.

The same caution applies as with --listen: an agent with access to ComfyUI can run workflows on your machine. Give it the local connection only on a computer you are happy for it to use.

Using ComfyUI from another device, safely

By default ComfyUI listens only on your own computer, at 127.0.0.1, port 8188. To open it from a laptop or tablet on the same network, start it with --listen (all network interfaces) or --listen followed by a specific address, and choose a port with --port. HTTPS is available with --tls-keyfile and --tls-certfile.

Think before you do. Anyone who can reach that address can run workflows on your machine, and with Manager on, install custom nodes, which means running code. If you expose ComfyUI with --listen, keep Manager at the Standard or Strict security level, never forward the port to the internet, and use a VPN if you need it away from home. On a shared server, --disable-auto-launch stops it opening a browser on start.

One more question that comes up often: is there an official ComfyUI Docker image? No. ComfyUI does not publish one. The images on Docker Hub are community builds, unsupported by the ComfyUI team, so read their Dockerfile before you trust one with your models.

ComfyUI or something simpler?

ComfyUI rewards curiosity: once you understand a workflow, you can change any part of it, and new open models usually work there first. If you would rather type a prompt and get a picture with no wires at all, a hosted tool such as Google's Nano Banana 2.1 in the free Gemini app is the simpler route, at the cost of daily limits and uploading your images. Our Nano Banana 2.1 guide covers that side. For chat models rather than images, Ollama and LM Studio are the equivalent local tools; our comparison explains which to pick.

A quick way to decide:

• You want local images, privately, with no limits: ComfyUI, starting with Desktop.
• You want the newest features the day they land, on Windows: ComfyUI Portable.
• You are on Linux, a server, or unusual hardware: the manual install, or comfy-cli.
• Your computer has no suitable GPU: Comfy Cloud, or a hosted model.

ComfyUI looks like a control panel from a spaceship the first time, and almost everyone closes the window once before trying again. That is normal. Pick the download that matches your computer, let the first workflow tell you what it is missing, and change one thing at a time. Jake now makes his window posters and his refurbished-phone photos on a PC that used to collect dust, with no limits and nothing leaving the shop. Ethan's only request was that he back up his workflows before every update. Your first picture is ten minutes away.

Revision note. Written October 7, 2026, two days after ComfyUI v0.39.0. If an older guide sent you to clone ComfyUI Manager by hand and nothing appeared, you did nothing wrong; the steps simply changed.