Comfy Agent

Describe what you want right inside ComfyUI. It plans, builds, and runs workflows. Editing the canvas while you work, it crafts side by side with you.

​

What it can do

• Find templates, models, and nodes available in your ComfyUI environment.
• Build a workflow from a template and customize its prompts and settings.
• Change part of the current workflow without rebuilding the rest.
• Add, connect, configure, and remove nodes.
• Validate a workflow before running it.
• Inspect images and read video or audio properties.
• Run workflows and return their outputs.
• Keep working on the canvas while the Agent responds — you don’t have to wait for it to finish, and it works from what you currently have open.
• Create a skill in plain language, or ask the Agent to use one of Comfy’s public skills instead of only your own. See Agent skills.
• Open more than one chat at once. Each chat is its own session with its own history and context.

​

Where you can use it

Either way, Agent requests use Comfy services over the network. See Data and privacy. Comfy Agent is separate from Comfy MCP and Comfy CLI. MCP connects an external AI client to Comfy. The CLI is for terminal and script workflows. Comfy Agent works in the ComfyUI interface. To make your first image, follow Get started with Comfy Agent.

​

What it costs

Free initial tokens are now available for everyone to try several Comfy Agent chats. After that, Comfy Agent charges from the same Comfy Credits pool you already use for generations. You need a positive balance in Comfy Credits to keep running the Agent, but you don’t necessarily need a subscription plan. See the subscription and top-up plans for plan details.

​

Pricing per token

​

Coming soon on local

Once Comfy Agent reaches Comfy Desktop, it will work with your own GPU, models, and custom nodes, and walk you through setup. It will also go beyond what’s on Cloud today: building custom editing tools directly on the canvas for you, such as a one-off masking or compositing tool you didn’t have to install.

​

Choose which workflow to edit

Use the workflow selector in the Agent panel to choose the workflow to edit, then send your message. Switching to another tab while the Agent is responding does not redirect its edits. The Agent cannot create or switch tabs itself. To work elsewhere, open the other workflow, choose it in the Agent panel, and send a new message. The Agent can read another open or saved workflow for comparison without loading it into the current tab. For example:

Compare this workflow with my portrait workflow and use the same sampler settings here.

​

Build and edit workflows

For a new workflow, open a blank workflow tab and choose it in the Agent panel. The Agent searches for a matching template before building from individual nodes. A template includes a connected graph and an output node. For changes to an existing workflow, name the part you want changed and what should stay the same:

• “Change the prompt to a watercolor landscape. Keep the model and sampler settings.”
• “Add an upscaling stage after image generation.”
• “Switch this workflow to Flux using models installed here.”

The Agent builds in real time on the canvas while you keep editing the canvas.

​

Validate before running

The Agent checks node connections, model names, required values, and output nodes before a run. It separates workflow errors from inputs that only you can provide. If the workflow is valid but still needs an image, video, or other input, attach the file and ask the Agent to continue.

​

Work with media

The Agent can inspect the visible contents and dimensions of a still image. For video, it can read the duration, dimensions, frame rate, and format. For audio, it can read the duration, format, and audio tracks. Use Attach a file in the message box to add an image or file. You can also drag media into the conversation. Media editing requires the media tools to be available in your installation. In Cloud, edited files are saved to your asset library. The Agent returns a preview or file link in the conversation.

​

Run permissions

Choose a run permission next to the message box:

• Ask requests approval before each workflow run.
• Auto runs workflows without asking for each run.

Select Save changes to apply your choice. In Ask mode, select Run in the approval message to start the workflow, or Cancel to decline that run. Ask and Auto control workflow runs. The Agent can edit the selected workflow in either mode.

​

What the Agent can use

Normally in a chat session, the Agent uses your message, selected workflow, selected nodes, attachments, and workflow errors. It can also read another workflow when you ask it to compare or reuse settings. Besides, the agent also has access to the saved workflows, available models and available nodes in the current workspace, as well as public template workflows and public/private skills, which will all be used when needed. For Cloud and local data flow, chat history, and privacy controls, see Comfy Agent data and privacy.

​

Stop or cancel

Select Stop in the message box to stop the Agent’s current response. Workflow execution has separate controls in the run and queue area; use those to manage a workflow that has already been submitted.

​

Next steps

​

Feedback

Comfy Agent is currently in beta. We are actively developing new features and fixing issues. Tell us what works and what doesn’t and it is the best way to help us prioritize the right things to build.

• Feedback: https://links.comfy.org/agentfeedback
• Discord: join our server and find us in #comfy-agent channel