# Comfy Local MCP

Windows-native MCP server for a local ComfyUI portable package. It talks to
ComfyUI over its loopback HTTP API and runs with the portable package's own
`python_embeded\python.exe`; no virtual environment is required.

## Portable layout

The MCP package is expected to be placed next to the portable package folders:

```text
ComfyUI_windows_portable\
  ComfyUI\
  python_embeded\
  comfy-local-mcp\
```

The installer also accepts an explicit `-PortableRoot` when the MCP package is
stored elsewhere.

## Install

From PowerShell, run `install.ps1` in the extracted package directory:

```powershell
.\install.ps1 -PortableRoot "F:\ComfyUI_windows_portable" -RegisterCodex
```

The installer:

- uses `ComfyUI\..\python_embeded\python.exe` directly;
- installs the pinned MCP dependency into this package's `vendor` directory;
- copies bundled API-format workflows into `ComfyUI\user\default\workflows`;
- uses `ComfyUI\output` as the output directory;
- optionally registers the MCP server with Codex.

No files are installed into ComfyUI's `python_embeded\Lib\site-packages`.

## Runtime configuration

- `COMFYUI_URL`: loopback ComfyUI URL, default `http://127.0.0.1:8188`.
- `COMFY_PORTABLE_ROOT`: portable root containing sibling `ComfyUI` and `python_embeded` directories.
- `COMFYUI_ROOT`: optional explicit ComfyUI directory.
- `COMFY_MCP_WORKFLOW_ROOT`: defaults to `ComfyUI\user\default\workflows`.
- `COMFY_MCP_OUTPUT_ROOT`: defaults to `ComfyUI\output`.
- `COMFY_MCP_ALLOWED_ROOTS`: semicolon-separated workflow/output roots.

The MCP tools `list_workflows` and `save_workflow` operate in the configured
workflow directory. `validate_workflow` and `run_workflow` accept either an
absolute path inside that directory or a workflow filename.

## Build a Release

From the repository root:

```powershell
.\tools\comfy_local_mcp\build_release.ps1 -Version "0.1.0"
```

The generated archive is written to `dist\comfy-local-mcp-0.1.0.zip`.

## Notes

ComfyUI must be running before calling `server_info`, `validate_workflow`, or
`run_workflow`. Models and custom nodes remain customer-local; the MCP package
does not download or share model weights.