Official Local Comfy MCP Integration

This project supports two complementary interfaces under one Agent Skill:

Agent request
  -> ComfyUI Skill application layer
      -> comfyui-skill CLI for known workflows and repeated execution
      -> official local Comfy MCP for live discovery, validation, and orchestration

The CLI remains the primary path for workflows already imported into this project. It exposes a stable business parameter schema and has low overhead for chat execution, scripts, and batch jobs.

The official local Comfy MCP is used when the Agent needs current ecosystem knowledge: finding templates, inspecting installed nodes and models, validating a raw workflow, adapting template slots, managing a submitted job, or operating the local ComfyUI lifecycle.

Connect The Official MCP

Register the server using Comfy’s local MCP connection guide. When the Agent host exposes the MCP tools, the Skill calls them directly.

Shell-only Agent hosts can use the repository bridge:

python ./scripts/comfy_mcp.py probe
python ./scripts/comfy_mcp.py tools
python ./scripts/comfy_mcp.py call search_templates --arguments '{"query":"background removal"}'

The bridge launches the official comfy-mcp stdio server. It does not duplicate or replace official MCP behavior.

Routing Rules

Request	Route
Run an existing `<server>/<workflow>`	`comfyui-skill` CLI
Repeat or batch a stable workflow	`comfyui-skill` CLI
Discover a current template, node, or model	Official MCP
Validate or adapt a raw workflow	Official MCP
Poll an MCP-submitted job and fetch its files	Official MCP
Install, update, restart, switch versions, or spend credits	Official MCP after explicit confirmation

For long MCP work, use run_workflow(wait=false) or run_template(wait=false), then job, then fetch_outputs. Once a route returns a prompt_id, stay on that route so one user request cannot accidentally generate twice.

The detailed Agent contract lives in references/comfy-mcp.md.