Comfy MCP: Turn Your AI Agent Into a Creative Technologist
Comfy Org launches Comfy MCP, connecting AI agents to ComfyUI via the Model Context Protocol. Agents can now search models, execute workflows, and generate images, video, and 3D content in natural language.
On June 30, 2026, Comfy Org announced Comfy MCP, connecting AI agents to ComfyUI's generation capabilities via the Model Context Protocol (MCP). The announcement came via the ComfyUI Newsletter, marking a major step toward natural language driven creative workflows.
With Comfy MCP, your agent can search models, nodes, and template workflows; read and execute shared workflow URLs; and generate images, video, 3D, and audio — all through natural language. No node graphs, no manual downloads, no local GPU required.
"Everything is now in natural language. No nodes, no download, no GPU, no node graphs if you don't want them." — Comfy Org
Three Ways to Connect
Comfy MCP ships as three complementary tools, each designed for a different use case:
1. Comfy Cloud MCP (Public Beta)
The hosted MCP server at cloud.comfy.org. Connects to Comfy Cloud — runs workflows on Cloud GPUs, uses pre-installed models, and searches the full template library. Best for chat-driven agents.
- Clients: Claude Desktop (OAuth sign-in via Custom Connector), Claude Code (install the
comfy-cloudplugin from Comfy Skills, or connect directly) - Headless/CI: Use an API key (
X-API-Keyheader) instead of browser OAuth - Beta feedback: Discord #comfy-mcp-and-cli or the feedback survey
Quick start for Claude Desktop: add a Custom Connector with URL https://cloud.comfy.org/mcp, sign in via OAuth, and start creating.
Quick start for Claude Code: register the Comfy Skills marketplace (/plugin marketplace add Comfy-Org/comfy-skills), install comfy-cloud (/plugin install comfy-cloud@comfy-skills), then run /mcp → comfy-cloud → Authenticate.
2. Comfy Partner MCP (Private Preview)
A local MCP server that provides unified generation tools across 30+ partner providers (BFL, Ideogram, Kling, Runway, Veo, Meshy, ElevenLabs, and more). Runs on your machine, connects to the Comfy API with a single API key. No local GPU required — all generation is API-based.
10 tools are available: generate_image, generate_video, generate_3d, generate_svg, generate_audio, generate_music, media_upload, read_media, models_explore, and balance.
- Requirements: Node.js 20+, pnpm 10, and a Comfy API key (
comfyui-...) - Clients: Claude Desktop, Claude Code, and Amp — add via MCP JSON config
- Status: Private preview; access is limited to invited partners
See the Partner MCP documentation for installation, configuration, and available tools.
3. Comfy CLI
The command-line companion for terminal workflows, scripts, and CI pipelines. Use comfy generate for one-shot partner API calls, or comfy run for full workflow execution on Comfy Cloud.
- Best for: batch jobs, automation, CI pipelines, shell scripts
- Auth:
comfy cloud login(OAuth) or API key (CI) - Output: save to disk (
--download) or JSON (--json)
What MCP Unlocks
The MCP protocol gives agents access to all the latest image, video, 3D, and audio models, plus hundreds of popular ComfyUI workflows that are automatically updated to best practices. This means your agent always knows what is up to date without manual workflow maintenance.
Example use cases from the announcement:
- "Take one hero sneaker shot and generate 20 ad variants in 4 aspect ratios for Meta Advantage+."
- "Read my saved product-shot workflow and re-run it on these 5 new images."
- "Turn three storyboard frames into a cinematic clip — then explain the workflow so my teammate can reuse it."
- "Generate a consistent character across 12 card-art poses."
Community MCP Servers for Local ComfyUI
For users who prefer to run ComfyUI locally, several community-maintained MCP servers already connect MCP clients to a local instance:
| Project | Repository |
|---|---|
| comfyui-mcp | artokun/comfyui-mcp |
| comfyui-mcp | shawnrushefsky/comfyui-mcp |
| comfyui-mcp-server | joenorton/comfyui-mcp-server |
| comfy-mcp-server | lalanikarim/comfy-mcp-server |
These are community-maintained — support, features, and compatibility vary by project. Setup for most takes 2-3 minutes:
git clone <repo-url> && cd <repo> && pip install -r requirements.txt && python server.py