Home/ COMFYUI/ Comfy Agent Tutorial: Image & Video Workflows in Comfy Cloud
COMFYUI

Comfy Agent Tutorial: Image & Video Workflows in Comfy Cloud

Learn how to use Comfy Agent in Comfy Cloud to build and refine image and video workflows with natural language, plus Agent vs MCP and local availability.

Published Oct 19, 2026 · 4 min read
Conceptual illustration of a chat bubble connecting to image-generation nodes and a product-video filmstrip

Updated October 18, 2026. Comfy Agent lets you describe a creative task in plain language and watch an AI assistant build, edit, and run the underlying ComfyUI node graph. This documentation-based guide separates current features from announced ones. The Signal has not independently tested Comfy Agent; the prompts below are suggested exercises, not recorded results.

What is Comfy Agent?

Comfy Agent is the built-in assistant for ComfyUI. Unlike a chatbot that merely suggests nodes, it can inspect an existing graph, modify connections, explain errors, and request permission to run generations. As of October 18, 2026, it is available in Comfy Cloud. Desktop/local integration is announced but not yet generally available according to the official product page.

How to get started in Comfy Cloud

  1. Open Comfy Cloud and sign in.
  2. Open an existing workflow or start with a template.
  3. Open the Agent panel and use its workflow selector to choose the workflow to edit. Switching editor tabs alone does not redirect the agent.
  4. Describe the output and constraints: subject, format, references, model preferences, and what should remain unchanged.
  5. Choose Ask in the run-permission control and save the change. Check the selected model, connections, required inputs, and output settings before approving Run. Auto mode allows runs without approval for each generation.
  6. Review the result and ask for one specific revision at a time. Save the workflow so it can be reused.

Example: build a product-image workflow

Try this prompt: “Create a reusable product photography workflow for a skincare bottle. Use the uploaded bottle photo as a reference, preserve its label and proportions, generate three neutral studio lighting variants, and keep the workflow editable. Show me the graph before running.”

This is a suggested exercise, not a benchmark we have independently executed. Results depend on the models and nodes available in your Cloud account.

Example: turn images into a short video

Try: “Starting from the selected product image, design an image-to-video workflow for a short vertical social ad. Preserve the product identity, use subtle camera movement, and show which video model and output settings you choose before generation.” Confirm model availability and credit use before running.

Comfy Agent vs Comfy MCP

FeatureComfy AgentComfy MCP
InterfaceBuilt into the ComfyUI canvasYour external AI assistant or coding agent
CloudAvailable nowAvailable now
Local ComfyUIComing soonAvailable through local MCP connection
Best forInteractive graph creation and iterationAgent-led automation from tools such as Codex

For programmatic queues, API calls, and n8n pipelines, see our ComfyUI automation workflows guide. If you are choosing between hosted GPUs and your own machine, read Comfy Cloud vs local GPU.

Pricing and limitations

Comfy Agent chat consumes Comfy Credits for language-model tokens, alongside applicable generation charges. The documentation offers initial trial tokens; after those, a positive credit balance is needed for chat, but a subscription is not necessarily required for the chat itself. This does not guarantee access to every Cloud generation service. Check your plan and selected model, the Agent billing documentation, and current Comfy pricing before running. Allowances and rates may change. The local version, local custom-node support within Agent, and local LLM options are roadmap items, not features to promise today.

Troubleshooting checklist

  • If the agent picks the wrong model, specify your available models and the desired task.
  • If outputs drift from the reference, ask it to preserve explicit attributes and inspect the reference inputs.
  • If a node fails, share the actual error and request a graph-level explanation rather than rerunning blindly.
  • For generation on your own hardware today, consider local Comfy MCP with a compatible client and local models. Your external assistant may still use a hosted LLM, and partner nodes may use paid services. Local MCP does not automatically mean fully offline operation.

Official references

For workflow selection, validation, and Ask/Auto controls, see the current Comfy Agent documentation. Availability was checked on October 18, 2026.

ComfyUI launch announcement (October 1, 2026) · Comfy Agent product page · Comfy MCP documentation.

The Signal newsletter

Keep getting this

One edition a week on open models, local setups and the tools around them.

Read the latest issue

Email delivery opens once the newsletter platform is connected.