On October 1, 2026, Comfy Org launched Comfy Agent, an AI assistant that lives inside the ComfyUI canvas and builds, runs and fixes node workflows from a plain English description. If you have ever spent an evening hunting for the one missing node in a 60-node graph, the pitch is obvious. But the launch also left a lot of people confused: is it free, does it work on my local ComfyUI, how much do credits cost, and how is it different from Comfy MCP? This Comfy Agent guide answers all of that, shows how to get started, gives you prompts that work well, and is honest about what the beta cannot do yet.
- Comfy Agent is a built-in AI agent that plans, builds, runs and debugs ComfyUI workflows directly on your canvas. It is in beta.
- Right now it only works on Comfy Cloud (cloud.comfy.org). The Desktop and local GPU version is promised “in a few weeks” but is not out yet.
- You get some free chats to start. After that, agent chat tokens and GPU time both draw from your Comfy Credits.
- It runs on Anthropic models. Choosing other LLMs, including local ones, is on the roadmap.
- Comfy MCP is a different product: it connects your own external agent (Claude, Codex, Cursor) to ComfyUI.
What Is Comfy Agent?
Comfy Agent is Comfy Org’s own AI agent, built into the ComfyUI app. You describe what you want in a chat panel next to the canvas, and the agent plans a workflow, places and connects the nodes, runs the graph and fixes errors it runs into. Comfy’s announcement puts it simply: “Describe what you want, and Comfy Agent can build, run, and iterate on workflows right in front of you!”
The important design choice is that it edits the canvas you already have open. It does not hand you a JSON file to import. You can watch nodes appear, stop it, tweak a parameter yourself, and carry on. Comfy calls this “vibe crafting”: the agent builds the graph in place while the workflow stays open and editable.
The launch post was written by Jo Zhang, Christian Byrne and Kishore Shimikeri, and Comfy describes the agent as knowing the latest models, ComfyUI extensions and workflows. Under the hood it runs on frontier models from Anthropic, with Comfy adding your workflow context, your error messages and its own ComfyUI knowledge to each request.
What Comfy Agent Can Do
Based on Comfy’s announcement and the early hands-on write-ups, the beta can:
- Build a workflow from a description, including starting from a template and customising prompts and settings.
- Edit part of a graph: add, remove, connect and configure nodes without rebuilding everything.
- Run the graph and check the result, then iterate. With your permission by default, or automatically in Auto mode.
- Explain an existing workflow, point out slow parts and trace errors back to the node that caused them.
- See what you see: it can interpret images you attach or view, and information from video and audio files.
- Take references: you can mention a node by name, drag in a library asset, or point it at another saved workflow.
- Run up to five chats in parallel, each with its own full history and context.
- Use skills: saved, reusable instructions such as your standard resolution or output format.
Comfy’s own example requests give a good feel for the intended use: “compare three more video models”, “explain this workflow’s bottlenecks”, and “rerun this product-shot workflow on 20 images”. These are exactly the boring, fiddly jobs that eat time in ComfyUI.
How to Start Using Comfy Agent
- Log in at cloud.comfy.org. You need a Comfy Cloud account. You do not need a local GPU.
- Open the Agent panel and claim your free chats. Comfy has not published the exact number of free chats.
- Pick a target workflow. Either open a saved workflow, or choose a blank workflow in the Agent panel.
- Ask for something small first. For example: “Build a text-to-image graph with an output node at 1024×1024.” Check the connections and required inputs before you let it run.
- Supply anything missing. If the graph needs a model or file the cloud does not have, you have to provide or replace it. Natural language cannot conjure up a missing checkpoint.
One detail that trips people up: the workflow selector decides which tab the agent edits for each request. Switching tabs while it is replying does not redirect it, and the agent cannot open or switch tabs by itself. If it seems to be “ignoring” your current tab, check the selector.
Ask vs Auto: The Run Controls Explained
| Mode | Graph edits | Running workflows | Use it when |
|---|---|---|---|
| Ask (default) | Allowed | You approve each execution | You are learning, or each run costs real GPU time |
| Auto | Allowed | Runs without asking and iterates until the result is ready | Batch jobs and tasks with a clear finish line |
The Stop button ends the agent’s current reply. It does not cancel a workflow that has already been submitted; that is handled by the normal run and queue controls. So if you hit Stop in Auto mode, also check the queue, or the GPU keeps working and keeps charging credits.
Agent Skills: Saving Your House Style
Skills are saved instructions with a name and a short description of when to use them. A skill can pin down things like default dimensions, background colour, which output node to use, or “always add a preview node before save”. Skills load by name or when a request matches their description, and edits to a skill apply from your next message.
There are personal skills and public preset skills, and you can paste in a SKILL.md file to turn it into a personal skill. Note that skills you have installed in an external coding assistant, such as Claude Code, are not automatically installed in Comfy Agent.
A good first skill for most people is a “house settings” skill: your usual aspect ratios, your preferred sampler and steps for the models you use, and your file naming pattern. It saves tokens too, because you stop re-explaining the same brief in every chat.
Comfy Agent Pricing and Credits
Comfy says the agent “is free to try and uses the Comfy Credits you already have.” In practice that means:
- New accounts get some free chats. The exact allowance is not published.
- After that, chat tokens draw from your Comfy Credits, the same pool that pays for Cloud GPU time. A subscription is not strictly required for chat, but you need a positive credit balance.
- Generation is billed separately as GPU time while workflows run.
Digital Production checked the agent’s token rates on October 7, 2026. With Opus 4.8 as the model, they listed these rates per 1,000 tokens:
| Token type | Credits per 1,000 tokens |
|---|---|
| Input | 1.055 |
| Output | 5.275 |
| 5-minute cache write | 1.319 |
| 1-hour cache write | 2.110 |
| Cache hit | 0.106 |
And the Comfy Cloud plans as listed in the same article:
| Plan | Monthly | Yearly | Job time limit |
|---|---|---|---|
| Standard | US$20 | US$192 | 30 minutes |
| Creator | US$35 | US$336 | 30 minutes (adds model and LoRA imports) |
| Pro | US$100 | US$960 | 1 hour |
Cloud jobs run on an RTX PRO 6000 Blackwell with 96 GB of VRAM, so memory is rarely the problem there. Prices and rates change, so check comfy.org/pricing before you commit.
How to keep costs down: write short, specific requests; save repeated instructions as a skill instead of retyping them; stay in Ask mode until you trust a workflow; and remember that long back-and-forth explanations cost tokens even when no image is generated.
Does Comfy Agent Work With Local ComfyUI?
Not yet. This is the most searched question about Comfy Agent, so to be clear: as of late October 2026, the agent works on Comfy Cloud only. Comfy said at launch that a Comfy Desktop version is coming “in a few weeks”, and that local Comfy will become “fully agentic” with guided setup, model suggestions based on your hardware, and help with missing dependencies. No firm date has been given.
Even when the Desktop version arrives, it will not be fully offline. Digital Production notes that on the Desktop route your files and chat history stay local, but your messages and relevant workflow information still pass through Comfy’s online service, because the language model runs there. Local nodes will use your GPU; partner nodes will call their providers.
If you want an AI assistant driving your local ComfyUI today, your options are Comfy MCP (next section) or a third-party community tool. For local heavy lifting on a 16 GB card, my 16 GB VRAM fixes for MiniMax H3 still apply no matter who builds the graph.
Comfy Agent vs Comfy MCP
These two names get mixed up constantly. Comfy MCP launched in public beta on June 29, 2026, and it is a different product.
| Comfy Agent | Comfy MCP | |
|---|---|---|
| What it is | An agent built into the Comfy app | A server that gives your own agent ComfyUI tools |
| Who does the thinking | Comfy’s agent (Anthropic models) | Your chosen client: Claude, Codex, Cursor, Hermes and others |
| Where it runs | Cloud GPUs now, local later | Comfy Cloud, or a local GPU |
| Custom nodes | Cloud-supported nodes only, for now | Cloud-supported nodes on Cloud; any node locally |
| Works on the canvas with you | Yes, live | No, it submits workflows and fetches outputs |
| Best for | Artists who want help inside ComfyUI | Developers who already live in a coding agent |
To set up Comfy MCP, Comfy suggests creating a Comfy account and then either adding it as a custom connector in Claude Desktop, or asking your agent to install it from the docs at docs.comfy.org/agent-tools/cloud. Comfy also has a beta Comfy CLI and a public comfy-skills repository on GitHub if you want to script things.
Example Prompts That Work Well
The agent works best with concrete goals and a clear finish line. These are good patterns to copy:
- Build: “Build a text-to-image workflow with FLUX 3 Image Klein distilled, 16 steps, 1024×1024, with a preview node and a save node named
klein_test.” (If you have not met Klein yet, see my FLUX 3 Image Klein ComfyUI guide.) - Explain: “Explain what this workflow does, node by node, and tell me which three nodes take the most time.”
- Debug: “This graph fails at the VAE decode. Find the cause and fix it without changing the model.”
- Compare: “Duplicate this image-to-video graph for two more video models and lay the outputs side by side.”
- Batch: “Run this product-shot workflow on the 20 images in my library folder
shoesand save each with the original file name.” - Refactor: “Group the prompt nodes, collapse the loaders, and add notes explaining each section for a beginner.”
Prompts that tend to go badly are vague ones (“make it look cinematic”), ones that depend on a model the cloud does not have, and ones that ask it to manage tabs.
Limitations and Gotchas
- It is a beta. Comfy says so plainly. Reliability across the huge range of custom nodes is unproven.
- Cloud only. No local GPU, local models or arbitrary custom nodes until the Desktop version ships.
- Cloud-supported nodes only. If your favourite node pack is not on Comfy Cloud, the agent cannot use it.
- Five chats is not five GPUs. You can run five chats at once, but generations still wait in the queue.
- Stop does not stop everything. Already-submitted jobs keep running until you cancel them in the queue.
- Talking costs credits. Every message spends tokens, even if no workflow runs.
- Cloud and local histories will not sync. Per Digital Production, chats on Cloud and on the future Desktop version are separate.
Privacy: What Leaves Your Machine
On Comfy Cloud, conversations, workflows and assets are stored on Comfy’s servers. On the future Desktop version, local files and chat history are expected to stay local, but messages and the workflow context the agent needs still travel through Comfy’s service to the language model. If you work with client material under NDA, factor that in, and keep sensitive assets out of the chat until you have read Comfy’s terms.
Roadmap
Comfy’s stated plans for the agent include:
- Comfy Desktop support, with local GPU, local models and any custom node.
- Guided setup and model suggestions based on your hardware.
- Agent-built custom nodes from a natural-language description.
- More model options, including choosing your LLM, with local LLM support on the list.
- Better public skills and built-in knowledge, and more external integrations.
Feedback goes through a form linked in the announcement and the #comfy-agent channel on Comfy’s Discord.
Frequently Asked Questions
Is Comfy Agent free?
It is free to try. New Comfy Cloud accounts get some free chats, and after that the agent uses Comfy Credits for chat tokens, with GPU time billed separately.
Can I use Comfy Agent with my local ComfyUI install?
Not yet. It works on Comfy Cloud only. Comfy has promised a Desktop version with local GPU support but has not given a date.
Which AI model does Comfy Agent use?
Comfy says it runs on frontier models from Anthropic. Digital Production’s pricing check listed Opus 4.8 rates. Choosing other LLMs is on the roadmap.
Does Comfy Agent support custom nodes?
On Cloud it can use nodes that Comfy Cloud supports. Arbitrary local custom nodes will need the Desktop version.
What is the difference between Comfy Agent and Comfy MCP?
Comfy Agent is built into ComfyUI and edits your canvas live. Comfy MCP lets your own external agent, such as Claude or Cursor, submit workflows to Comfy Cloud or a local ComfyUI.
Will Comfy Agent work offline?
No. Even on Desktop, your messages and workflow context go through Comfy’s online service, because the language model runs there.
How many chats can I run at once?
Up to five parallel chats, each with its own history. Generations from those chats still share the GPU queue.
Final Thoughts
Comfy Agent is the biggest change to how people use ComfyUI since the node editor itself. It will not replace knowing what a VAE or a sampler does, and the beta is cloud-only with credit costs to watch. But for explaining someone else’s workflow, fixing a broken graph, or doing the same job 20 times, it already saves real time. Try it with the free chats on a small task, keep Ask mode on, and save your house settings as a skill early.
If you want ready-made graphs to give the agent as a starting point, try these:
- Best MiniMax H3 Text-to-Video Workflow for ComfyUI
- MiniMax H3 RIFE Frame Interpolation Workflow
- ComfyUI Pixal3D Workflow: Image to 3D
- MiniMax H3 Prompt Guide
