Comfy Vault

PyPI CI

An MCP server that exposes comfy-vault to coding agents, so you can ask Claude or Codex to audit and clean up your ComfyUI model library in plain language.

It wraps the comfy-vault library (PyPI package comfy-model-vault, a real dependency — not vendored) as five tools. The agent calls them; comfy-vault does the work.

Tools

Tool	What it does
vault_scan	Index a models directory recursively; counts, sizes, categories.
vault_orphans	List models no workflow references (read-only).
vault_lora_info	Look up Civitai metadata for a file by its content hash.
vault_duplicates	Find byte-identical duplicate models (same SHA256).
vault_delete	Delete orphans — dry-run unless confirm=true, fenced to the scanned roots.

Install

pip install comfy-vault-mcp

This installs comfy-model-vault (the comfy-vault library) as a dependency. Python 3.10+.

Use with Claude Code

claude mcp add comfy-vault -- comfy-vault-mcp

or add it to .mcp.json in your project:

{
  "mcpServers": {
    "comfy-vault": {
      "command": "comfy-vault-mcp"
    }
  }
}

Use with Codex CLI

Add to ~/.codex/config.toml:

[mcp_servers.comfy-vault]
command = "comfy-vault-mcp"

Example

Once connected, ask in natural language:

"Find all unused Pony LoRAs over 2 GB in D:\ComfyUI\models, check which of them still exist on Civitai, and show me the duplicates — but don't delete anything yet."

The agent will call vault_orphans, vault_lora_info, and vault_duplicates, then summarise. vault_delete stays a dry-run until you explicitly confirm.

Offline / testing mode

Set COMFY_VAULT_CIVITAI_FIXTURES to a directory of recorded {sha256}.json Civitai responses, and vault_lora_info resolves from those instead of the live API — no network, no key. This is how the test suite runs. For live use, an optional CIVITAI_API_KEY raises your rate limit but is not required.

License

MIT. Same license as comfy-vault, for consistency across the toolkit.