Tool Pick | comfy-mcp — Comfy's Official MCP Server That Lets Agents Run ComfyUI on Your Machine

🌏 中文版

Tool Info

What Problem It Solves

Ever tried using an agent for image or video generation tasks, only to find it can merely tell you which ComfyUI workflow and nodes to use — while installing packages, launching the server, running the workflow, and checking outputs are all on you? ComfyUI's native workflow is drag-and-drop nodes in a browser and click Queue. Agents had no standard way to touch the ComfyUI instance running on your machine.

comfy-mcp is a local MCP server maintained by Comfy-Org (the company behind ComfyUI). It wraps the entire comfy-cli into 39 MCP tools for agents to call: from run_workflow and generate_image for direct execution, to validate_workflow and workflow_deps for pre-flight checks on whether the required node packages are installed, search_templates / fetch_template for pulling ready-made workflows from the official template library, and even launch_comfyui / install_node so the agent can start the server and install packages on your behalf. Every tool calls comfy --json --where local under the hood — essentially turning commands you'd type manually in the terminal into structured, agent-callable interfaces.

Best-fit scenarios: automating image/video generation pipelines with Claude Code or Cursor — batch-running workflow variants, having the agent debug "why won't this workflow produce output" (via validate_workflow / node_dependencies to find missing nodes), or letting the agent pick a template from the official library and tweak parameters to your requirements.

Quick Start

Installation

pip install comfy-mcp "comfy-cli>=1.14.0"
comfy install    # if you don't have a ComfyUI workspace yet
comfy launch     # start ComfyUI

Basic Usage

// mcp.json for Claude Desktop / Cursor
{
  "mcpServers": {
    "comfy-mcp": {
      "command": "comfy-mcp",
      "env": {
        "COMFY_BIN": "/path/to/venv/bin/comfy"
      }
    }
  }
}

Core tools available once the agent connects:

• server_info() — check if ComfyUI is running and what hardware is available
• search_templates(query, tag, model) — search the official template library for workflows
• run_workflow(workflow_path, wait, confirm_spend) — execute a workflow
• job(action="status|wait|watch|cancel", prompt_id) — monitor queued jobs
• fetch_outputs(prompt_id, out_dir) — retrieve generated images/videos

Advanced Usage

# Validate that all node packages a workflow needs are installed before running
comfy-mcp validate_workflow --workflow_path ./my_workflow.json

# Check dependencies and install missing nodes
comfy-mcp workflow_deps --workflow_path ./my_workflow.json

You can also fan out a single workflow into multiple parameter variants for the agent to batch-compare:

comfy-mcp vary_workflow \
  --workflow_path ./portrait.json \
  --slots '{"steps": [20, 30, 40], "cfg": [4, 7]}' \
  --out_dir ./variants

Comparison with Alternatives

Caveats

• Dual-licensed under AGPL-3.0-or-later or a commercial license — not plain MIT/Apache. If you're integrating comfy-mcp into a commercial product offered as a service, AGPL's copyleft clause may require you to open-source the combined work. Check whether your use case qualifies, or negotiate a commercial license.
• Still in Beta — the README itself labels it "Status: Beta." CI currently runs only on Python 3.10 and 3.14, with no explicit guarantee for versions in between.
• Spend-guarded tools: tools like partner_generate that call paid partner APIs always prompt for confirmation. run_template / run_workflow require an explicit confirm_spend=True to trigger the confirmation prompt, preventing agents from accidentally burning through your API credits.

Takeaway

What makes comfy-mcp interesting isn't just "ComfyUI got an MCP interface." It's that the tool took the operational logic originally designed for humans typing commands in a terminal via comfy-cli and brought the full set — all 39 tools — into the agent-callable world. That includes steps like "validate dependencies before running" and "search the template library" — pre-flight checks that an experienced user would know to do, now structured as actions the agent can proactively call. The lesson: when a mature CLI tool connects to MCP, the real value isn't wrapping a thin adapter layer — it's also structuring those "extra steps an experienced user would take" into callable tools.

References

• Comfy-Org/comfy-mcp — GitHub
• Open Sourcing Comfy MCP on Local — Comfy Blog