Comfy Agent Arrives: AI Builds and Repairs Image Workflows

Comfy, the company behind the node-based image generation platform ComfyUI, has launched Comfy Agent, an AI assistant that builds, edits, and repairs visual generation workflows inside Comfy Cloud. Announced on October 1, the release turns the canvas into a surface the agent can act on: the assistant searches for models, nodes, templates, and assets, assembles workflows, inspects errors, and executes runs.

The launch is part of a broader shift across creative and engineering software, where AI assistants are moving from describing work to manipulating the application's native objects. In Comfy's case, that means the agent can connect components and adjust settings while the user keeps full visibility of the canvas, inspecting node parameters and continuing to edit by hand.

According to a report by Ground Truth, the interface supports up to five parallel chats, letting users run several threads of workflow assistance at once. Comfy Agent makes the canvas an action surface for conversational assistance, rather than a separate tool the assistant can only talk about.

From conversation to canvas edits

The appeal of a visual workflow is control: models, processing nodes, templates, settings, and the connections between them are all visible and adjustable. That same detail makes setup and troubleshooting harder for someone who knows the image or video they want but not the exact graph needed to produce it. The agent is designed to bridge that gap, finding the right pieces and wiring them together.

The same feedback loop shows up across agent products: propose a change, execute it, inspect the result, and revise. A broken node or a failed run gives the agent a concrete error to work from, which is arguably a more forgiving environment for AI assistance than open-ended creative judgment. Ground Truth describes the agent's stated capabilities rather than an independent measure of how often it completes tasks.

The launch also carries a caveat familiar to agent products: an edit is not a recommendation. A generated workflow is an artifact the user can inspect and refine, and that inspection matters when an edit changes resource use or processing behavior. The company says users can continue editing the canvas and check every node parameter themselves.

Pricing, privacy, and the fine print

Comfy Agent is free to try, then draws from Comfy Credits, the same pool used for generations. Its documentation lists token charges, including a named model rate. Comfy says the agent is available to everyone in Comfy Cloud, with Desktop support promised in a few weeks.

The privacy documentation deserves attention from users drawn to local creative software. Cloud workflows run on Cloud graphics processors. In local use, chat history and workflow files stay on the machine and local nodes run on local hardware. However, the agent's messages and relevant workflow information are processed through Comfy's online service. Storage location, execution location, and reasoning location are three different questions, and the word local does not cover all three.

Permissions are split as well. Comfy documents approval and automatic modes for running workflows, but the agent can edit in either mode. A run approval is not a blanket checkpoint on every canvas modification, so reviewing the resulting graph is part of a sound workflow. This mirrors a wider trend across the AI News landscape, where agent permissions are getting granular treatment as the tools grow more capable.

The ecosystem follows the platform

The official launch is already reshaping the community around ComfyUI. The maintainers of comfyui-mcp, a popular community project that drove ComfyUI from any large language model over the Model Context Protocol, have announced the project is no longer maintained. According to the project's readme on GitHub, the repository will be archived on October 9, with its community Discord going read-only the same day. The maintainers point users toward Comfy's official agent and MCP tooling, describing deeper integration than a community effort could match.

Comfy has also introduced Comfy MCP in public beta, which the company calls the first MCP built for production pipelines. According to Comfy's blog post, it connects existing workflows to agents across Claude, Codex, Hermes, Cursor, and other MCP clients. It offers auto-updated best-practice workflows alongside access to the latest image, video, 3D, and audio models, all driven in natural language with no node downloads, no GPU setup, and no graph editing required. Workflows can be shared by URL and re-run by teammates' agents, with every generation designed to be fully reproducible.

The pattern is familiar: as brands and platforms map out what AI says about their products, as covered in this earlier report on agent-native brand intelligence, official tooling is consolidating around the platforms themselves. For creative workers who live inside node graphs, the arrival of a native agent marks a new phase: the tools that once only executed instructions are starting to build the pipelines themselves.

Sources: Ground Truth, Comfy blog, GitHub project documentation.