Skip to main content
The @image tool generates and edits images via the llm/imagegen package, local/keyless-first and multi-provider.

​
Backend selection (local first)

imagegen.NewFromEnv picks one, degrading to null when nothing is configured:
  1. CHATCLI_IMAGE_PROVIDER=sdwebui → local Stable Diffusion WebUI (/sdapi/v1/txt2img, default http://localhost:7860). Keyless — the preferred path.
  2. CHATCLI_IMAGE_URL → an OpenAI-compatible endpoint (/images/generations). Covers OpenAI, LocalAI and similar.
  3. OPENAI_API_KEY → OpenAI Images API (default gpt-image-2).
  4. CHATCLI_IMAGE_PROVIDER=responses (or openai + CHATCLI_IMAGE_API=responses) → OpenAI Responses API, where a chat model like gpt-5.6-sol generates the image via the image_generation tool.
  5. GOOGLEAI_API_KEY/GEMINI_API_KEY → native Google (Gemini image / “Nano Banana” :generateContent, generate and edit; default gemini-3.1-flash-image — the Imagen :predict family was shut down by Google on 2026-08-17).
  6. XAI_API_KEY → native xAI Grok Imagine (OpenAI shape, omits size; default grok-imagine-image-2.0). Generate only.
  7. CHATCLI_IMAGE_PROVIDER=bedrock → AWS Bedrock (Stability / Nova Canvas via InvokeModel, reusing the chat provider’s credential chain).
CHATCLI_IMAGE_PROVIDER pins the backend (sdwebui|url|openai|responses|google|xai|zai|minimax|bedrock); CHATCLI_IMAGE_API selects images|responses on OpenAI; CHATCLI_IMAGE_MODEL sets the model.
The model is inherited by both generation AND editing. @image gen and @image edit read the same CHATCLI_IMAGE_MODEL/CHATCLI_IMAGE_PROVIDER — one /model-image serves both. You never configure “one model to create and another to edit”.

​
The @image tool

edit accepts image (path), mask (optional PNG for inpainting), strength (0..1 — how much to change). The result reports the output file path(s).

​
Multi-provider editing (not locked to one)

Editing is a per-API capability — not every generation backend has an edit endpoint. @image edit is capability-gated: editing backends get Edit; generation-only ones refuse cleanly (and route, see below).

​
Inheritance and automatic routing

Editing inherits the active provider/model. Only when the active backend can’t edit does @image edit route to an edit-capable fallback — and says so explicitly:
ℹ️ Image backend "xai" can’t edit images; routed to "sdwebui" for this edit. Set CHATCLI_IMAGE_EDIT_PROVIDER to pin a different one.
Fallback order (each credential-gated, never routing to something that can’t run):
You can pin the editor without changing the generation provider:
If no editor is available, the tool returns a clear error listing the editing backends and how to configure them.

​
Models & per-command config

Catalog (summary, Aug 2026): OpenAI Images gpt-image-2 (default; the older gpt-image-1/1.5/1-mini are deprecated and shut down late 2026); OpenAI Responses gpt-5.6-sol/gpt-5.6-terra/gpt-5.6-luna, gpt-5.5; Google gemini-3.1-flash-image (default), gemini-3.1-flash-lite-image, gemini-3-pro-image (gemini-2.5-flash-image shuts down 2026-10-02; Imagen was shut down 2026-08-17); xAI grok-imagine-image-2.0 (default), grok-imagine-image; Z.AI glm-image, cogview-4-250304; MiniMax image-01; Bedrock stability.stable-image-core-v1:1 (default), stability.stable-image-ultra-v1:1, stability.sd3-5-large-v1:0, amazon.nova-canvas-v1:0 (legacy, EOL 2026-09-30).

​
Formats & details

  • OpenAI edits accepts image/jpeg, image/png, image/webp (not gif). The multipart part’s Content-Type is sent correctly from the image bytes.
  • SD WebUI img2img uses denoising_strength (mapped from strength); default 0.6 (visible change while preserving structure).
  • Bedrock Stability uses mode: image-to-image + strength; Nova Canvas (legacy) uses IMAGE_VARIATION with similarityStrength.

​
Privacy & cost

  • Local Stable Diffusion WebUI is keyless and generates and edits on your machine — ideal for privacy and zero cost.
  • Cloud backends use the user’s own key.
  • Confirm before generating/editing large batches; mention cost when relevant.