Comfy MCP Open-Sources Local AI Agent Control for Node Workflows

ComfyUI, the node-based AI image and video generation platform, open-sourced Comfy MCP, a Model Context Protocol server that lets AI agents interact directly with local ComfyUI installations. The release allows developer tools and assistants like Claude Desktop, Cursor, and Codex to inspect installed custom nodes, build image and video workflows, and execute runs directly on the creator's local hardware.

What's new

Comfy MCP acts as an open-source bridge using Anthropic's Model Context Protocol standard. Instead of manually wiring nodes in the web interface, creators can use conversational LLMs to generate graph structures, adjust node parameters, and run pipelines against local checkpoints, LoRAs, and custom extensions.

The server runs locally alongside ComfyUI, meaning prompt instructions, custom node dependencies, and rendered outputs stay on the user's machine. Key capabilities available to connected agents include:

• Node discovery: Reading the local environment to identify installed custom nodes, expected input types, and parameter boundaries.
• Workflow generation: Converting text prompts into valid ComfyUI JSON graph formats.
• Execution and debugging: Triggering generation jobs on local GPUs and reading error logs when a node pipeline fails.

How it fits your workflow

For technical directors, VFX artists, and pipeline developers, Comfy MCP provides a direct link between code editors and visual generation node graphs. Creators building automated pipelines can use coding environments like Cursor to write Python nodes or refactor massive JSON workflows without manually dragging connections in the browser interface.

Compared to cloud API agent frameworks like Replicate or Fal.ai, Comfy MCP keeps execution local, eliminating compute costs per run and protecting proprietary asset pipelines. While traditional user interfaces like Automatic1111 require manual input for every slider, Comfy MCP allows natural-language agents to modify intricate ControlNet, IP-Adapter, or AnimateDiff pipelines dynamically. It positions ComfyUI as a backend engine for AI-assisted programming environments rather than just a standalone graphical interface.

What it costs / how to try it

Comfy MCP is free and open-source under the MIT license. It requires a local installation of ComfyUI, a working local GPU setup, and a supported MCP client such as Claude Desktop or Cursor.