Agents and the API
ImageForge is an MCP server, so Claude, Cursor or any MCP client can generate images through it.
Open API and it prints the configuration to paste into your client. It names the engine's own interpreter and a working directory, so it works as-is once the engine is installed — before that it says so instead of handing you a configuration that cannot start.
Seven tools are exposed:
| Tool | Does |
|---|---|
generate_image |
Text to image |
edit_image |
Image to image |
inpaint_image |
Replace part of an image |
assist_prompt |
Improve a prompt |
list_models |
The catalogue, with each model's licence |
rate_output |
Record a preference |
generation_insights |
What has worked so far |
list_models carries licences deliberately: an agent asked for something
commercial can see that SDXL-Turbo forbids it.
There is also an HTTP API on 127.0.0.1:8765 with the same engine behind it.
The API screen lists every endpoint, shows ready-made curl and Python
snippets, can create a key, and has a console for trying a call without
leaving the app.
Both are local only. Nothing listens on the network.
Keys. Naming a file on disk in an API request requires the local key, even when keys are otherwise optional — otherwise any other program on your computer could ask ImageForge to read your images. Sending an image as base64 needs no key.
Folders. The MCP tools that take a file path will only open images inside ImageForge's own outputs, datasets and cache folders. An agent's instructions can come from whatever it last read, so the thing choosing that path may not be you. To work on photos kept elsewhere, set
IMAGEFORGE_IMAGE_DIRSto those folders — point it at a pictures folder, not at your whole home directory.