ComfyUI MCP server lets AI agents control workflows with plain-text prompts
Comfy.org launched a beta MCP server that lets AI agents run ComfyUI workflows from natural-language descriptions, search models and nodes, and rerun saved pipelines with new inputs.
Comfy.org has opened beta access to an MCP server that turns ComfyUI into an agent-controllable platform—users can now describe a workflow in natural language and have an AI execute it without manually wiring nodes.
The Model Context Protocol server exposes five core capabilities: running workflows from plain-text descriptions, searching the ComfyUI ecosystem for models and custom nodes, loading workflows by URL, rerunning saved pipelines with fresh inputs, and pulling from a library of hundreds of pre-built templates that update automatically. The documentation went live alongside the beta.
Comfy.org also released a beta CLI tool and a Comfy Skill repository for agent frameworks. The skill repo packages ComfyUI operations as reusable modules that agent orchestrators can call without custom integration work. Both tools are designed to lower the barrier for developers building agentic workflows around image and video generation—no manual node graph required.
