Comfy Agent is now available inside Comfy Cloud, letting creators describe an image or video task and build its workflow through chat. It can help edit and troubleshoot the node-based canvas, with product visuals and model comparisons among the practical uses. Local support is still coming soon, and continued Agent use consumes credits.
Table of contents
What Comfy Agent Does
ComfyUI uses a graph of connected nodes to build image and video workflows. Comfy Agent adds a chat-based way to work with that graph. It can explain a workflow, add or change nodes, suggest models and help fix errors. The canvas remains editable, so users can inspect the steps behind a result rather than receive only a finished file.
The launch examples include comparing several models with one prompt, placing a product reference against different backgrounds and combining workflows. For a creator, these are useful starting points for testing a visual direction before spending time on a full campaign. They are examples of tasks to request, not a promise that every generated result will be ready to use.
The Agent can read an existing workflow and work on a selected part without rebuilding everything. It checks connections, required inputs and output nodes before execution. Missing reference files still need to be supplied by the user. Its workflow controls also let users reference nodes, assets and other workflows with the @ menu.
Product Visuals and Repeatable Creator Work
For product visuals, the useful part is being able to keep the workflow while trying another background, model or processing step. A creator can review one change at a time and retain the setup for the next brief. Product shape, logos, text and colours still need careful checking against the original reference.
Comfy also supports personal and public skills. Personal skills save reusable instructions, such as a square canvas, a preferred background and a required output node. Give the skill a name and explain when it should apply; the Agent can use it when requested or when its description matches the task. Changes apply from the next message.
This is a different approach from a single-purpose generator. Readers exploring other creative AI tools can also look at Adobe Firefly’s music, speech and sound-effects tools. Comfy’s focus here is the editable workflow around the output.
How to Start on Comfy Cloud
Sign in to Comfy Cloud, open a blank workflow and choose Agent beside the avatar. Select the intended workflow in its panel, then describe the image and ask it to build and check the graph without running it.
The getting-started guide recommends reviewing the connected workflow and supplying any missing input before generation. Select Ask for run permissions, request a run and approve it when ready. After viewing the result, ask for a specific change while keeping other settings unchanged.
Ask mode requires approval for workflow runs; Auto allows runs without approval each time. Both modes allow the Agent to edit the chosen workflow. Also, stopping its chat response is separate from cancelling a generation already submitted to the queue.
Free Trial, Credits and Plan Prices
New users can try five GPU runs and receive initial Agent credits. This is not unlimited free generation. Agent chat uses Comfy Credits for language-model tokens, while workflows consume credits under the normal generation charges. There is no additional Agent surcharge on those workflow runs.
Comfy Agent is not restricted to one subscription tier. Free introductory credits or a positive credit balance provide access, and the Agent pauses when credits run out. A subscription is not necessarily required just to use it.
The current Comfy pricing page lists Standard at US$20 per month, Creator at US$35 and Pro at US$100. Annual billing is separate: US$192, US$336 and US$960 respectively. These are dollar prices, not announced India-specific rupee prices.
Local Support and Data Handling
Comfy Agent is available on Cloud now. Support for local ComfyUI in Comfy Desktop remains coming soon, with no confirmed release date. Planned local support includes working with the user’s GPU, models and custom nodes; it should not be treated as a feature available today.
Cloud stores conversations, workflows and assets, and processes messages through Comfy’s online service. The data-handling documentation also says the planned local setup sends messages and relevant workflow context through that service. Local GPU execution therefore does not mean an offline Agent.
For creators, the main reason to try Comfy Agent is easier workflow building with controls still visible on the canvas. Start with a small task, review the graph and output, then decide whether the saved setup is useful enough to repeat.