nc://gmi-image

GMI Image: pictures from GMI Cloud's Seedream, Gemini and the rest of its live catalog, as the gmi_image tool and /gmi-image. A Nodecode rewrite of zkasuran/hermes-gmi-media by zkasuran.

The original is an image backend for another agent, plugged into its one image_generate tool. Here the same wire is a cell of its own: one tool, one command, and the pictures written under a folder you name. Every call spends GMI credits.

The cell reads the model's own schema before it calls: GMI's catalog declares each model's parameters, their options and bounds, and which inputs it requires. The prompt goes to the parameter named prompt or text; the aspect goes to aspect_ratio when the model has one, and otherwise becomes the nearest of the sizes it offers, so a model can answer a shape other than the one asked for. The picture is downloaded as soon as the job finishes, because GMI's links may expire.

Get it

Tell the agent "add the gmi-image cell", or run nodecode add nc://gmi-image.

Set it up

You need a GMI Cloud API key from https://console.gmicloud.ai. Then either:

• put it in the environment variable GMI_API_KEY, which is what the cell reads unless the section names another variable, or
• name a file holding it in the section's key_file.

The section gmi_image in ~/.nodecode/config.jsonc is optional; config.example.jsonc shows it:

{
  "gmi_image": {
    "api": "https://console.gmicloud.ai/api/v1/ie/requestqueue/apikey",
    "key_env": "GMI_API_KEY",
    "model": "",
    "output_dir": "~/.nodecode/images"
  }
}

model left empty lets the cell probe its own preference list - seedream-5.0-lite, gemini-2.5-flash-image, gpt-image-2-generate - and take the first one the key can reach that accepts these inputs; name a model id to pin every call. output_dir is where pictures land (~/.nodecode/images by default), and timeout is how long one queued job may take. With no key set the cell still starts: the tool and the command answer with the setting that names one, and nothing is dialled.

Use it

Ask for a picture - "draw a blue boat on a calm lake, wide shot" - and the model calls gmi_image. The answer carries the file's path under output_dir, its size and its pixel size, and the model says where it is.

From the keyboard: /gmi-image a blue boat, /gmi-image --portrait the same boat at night --image https://example.com/boat.png, /gmi-image --model seedream-5.0-lite a heron.

From eval, the same call:

(gmi-image:generate-image "a blue boat on a calm lake" :aspect "landscape")

A failed call says the HTTP status or GMI's own job error, and the setting that usually fixes it; the key itself never appears in an answer. A job still running is polled until the section's timeout; a model that needs a source picture is only chosen when the call names one.

What the original does that this does not

Nothing of its request shape: the same requestqueue base, the same one-hour schema cache, the same mapping of a prompt and an aspect onto a model's declared parameters, the same aspect-to-nearest-size fallback under a 2048x2048 pixel ceiling, image inputs as public URLs only and at most three, the same submit-then-poll job with its growing delay and deadline, the same outcome media URLs, and the same scrubbing of the key out of every error. It is not a provider in another agent's catalog here; it is a cell with a tool of its own. See NOTICE.