ottoport 1.4.1 → 1.5.0

package/.claude-plugin/plugin.json

@@ -1,7 +1,7 @@
 {
   "name": "ottoport",
   "description": "One API for every model. Call chat, image, video, speech and music models through the OttoPort gateway — as MCP tools, slash commands, or the bundled CLI.",
-  "version": "1.4.1",
+  "version": "1.5.0",
   "author": {
     "name": "LITBOX LLC",
     "email": "support@ottoport.ai"

package/README.md

@@ -22,6 +22,7 @@ ottoport chat  "<prompt>" [--model claude-sonnet-5] [--system "..."] [--no-strea
 ottoport image "<prompt>" [--model gpt-image-2] [--size 1024x1024 | --resolution 2k]
                           [--aspect-ratio 16:9] [--n 1] [--seed 7]
                           [--image <url>] [--ref <url>[,<url>…]] [--out file.png]
+ottoport image --model qwen-image-layered --image <url> [--layers 4]
 ottoport video "<prompt>" [--model kling-3.0] [--duration 5] [--resolution 1080p]
                           [--aspect-ratio 16:9] [--image <url>] [--last-frame <url>]
                           [--ref <url>[,<url>…]] [--out clip.mp4]
@@ -86,6 +87,11 @@ first-and-last-frame (`image_url` + `last_frame_url`), and reference-to-video
 (`reference_image_urls`). Both take a `resolution`: `480p`…`4k` for video,
 `1k`/`2k`/`4k` for images, and a tier above a model's base rate costs more.
 
+Two image models take a picture apart instead: `qwen-image-layered` and
+`seedream-5.0-layers` split `image_url` into RGBA layers and return one URL per
+layer, bottom first (the prompt is an optional caption; `num_layers`, 2–10, is
+honoured by Qwen only). Each layer returned bills separately.
+
 Which modes and tiers a model accepts differs per model, so
 `ottoport_list_models` — and `ottoport models` in the CLI — prints each one's
 menu beside its price. For one model in full, including every parameter it

package/cli/ottoport.mjs

@@ -225,11 +225,13 @@ function references(flags) {
 }
 
 async function cmdImage(prompt, flags) {
-  if (!prompt) die("image requires a prompt: ottoport image \"a red fox\"");
+  // A layer-decomposition model needs only --image; the prompt is a caption.
+  if (!prompt && !last(flags.image)) die("image requires a prompt: ottoport image \"a red fox\" (or --image <url> on a layer model)");
   const { baseUrl, apiKey } = config(flags);
   const body = {
     model: last(flags.model) || "gpt-image-2",
-    prompt,
+    ...(prompt ? { prompt } : {}),
+    ...(last(flags.layers) ? { num_layers: Number(last(flags.layers)) } : {}),
     ...(last(flags.size) ? { size: last(flags.size) } : {}),
     ...(last(flags.resolution) ? { resolution: last(flags.resolution) } : {}),
     ...(last(flags["aspect-ratio"]) ? { aspect_ratio: flags["aspect-ratio"] } : {}),
@@ -246,7 +248,9 @@ async function cmdImage(prompt, flags) {
   if (!res.ok) die(await readError(res));
   const json = await res.json();
   const urls = (json.data ?? []).map((d) => d.url).filter(Boolean);
-  urls.forEach((u) => console.log(u));
+  // A layer stack names each line, so `headline` can be found among eight URLs.
+  (json.data ?? []).filter((d) => d.url).forEach((d) =>
+    console.log(d.layer ? `layer ${d.layer.index}${d.layer.name ? ` (${d.layer.name})` : ""}: ${d.url}` : d.url));
   if (last(flags.out) && urls[0]) await download(urls[0], last(flags.out));
 }
 
@@ -334,6 +338,8 @@ Usage:
   ottoport image "<prompt>" [--model gpt-image-2] [--size 1024x1024 | --resolution 2k]
                             [--aspect-ratio 16:9] [--n 1] [--seed 7]
                             [--image <url>] [--ref <url>[,<url>…]] [--out file.png]
+  ottoport image --model qwen-image-layered --image <url> [--layers 4]
+                            split one image into RGBA layers (prompt optional)
   ottoport video "<prompt>" [--model kling-3.0] [--duration 5] [--resolution 1080p]
                             [--aspect-ratio 16:9] [--seed 7] [--image <url>]
                             [--last-frame <url>] [--ref <url>[,<url>…]]
@@ -358,6 +364,7 @@ Examples:
   ottoport video "cinematic product reveal" --model seedance-2.0-fast --resolution 1080p
   ottoport video "the logo unfolds" --image start.png --last-frame end.png
   ottoport image "her, on a beach" --ref face.png,style.png
+  ottoport image --model seedream-5.0-layers --image https://…/poster.png
   ottoport speech "Welcome to OttoPort" --voice alloy --out welcome.mp3
   ottoport music "lo-fi focus beat" --duration 20 --out focus.mp3
   ottoport mcp

package/mcp/server.mjs

@@ -59,7 +59,10 @@ function accepts(capabilities) {
   if (Array.isArray(capabilities.modes) && capabilities.modes.length) {
     parts.push(capabilities.modes.join("/"));
   }
-  if (typeof capabilities.maxInputImages === "number") {
+  if (capabilities.output === "layers") {
+    // Not a reference image: the one input IS the job, and the answer is a stack.
+    parts.push(`one image_url, split into RGBA layers${capabilities.maxLayers ? ` (num_layers 2–${capabilities.maxLayers})` : " (count chosen by the model)"}`);
+  } else if (typeof capabilities.maxInputImages === "number") {
     parts.push(capabilities.maxInputImages > 0 ? `up to ${capabilities.maxInputImages} reference images` : "text only");
   }
   if (typeof capabilities.maxReferenceImages === "number") {
@@ -87,7 +90,10 @@ async function describeModel(id) {
     const refs = caps.maxReferenceImages ? ` (up to ${caps.maxReferenceImages} reference images)` : "";
     lines.push(`Modes: ${caps.modes.join(", ")}${refs}`);
   }
-  if (typeof caps?.maxInputImages === "number") {
+  if (caps?.output === "layers") {
+    lines.push("Output: the input image split into RGBA layers — one `data` entry per layer, bottom first, each with a `layer` object (index; name, z_index and bounding_box where the model reports them). Billed per layer returned.");
+    lines.push(`Input images: exactly one, in image_url. The prompt is an optional caption.${caps.maxLayers ? ` num_layers picks the count, 2–${caps.maxLayers}.` : " The model picks the layer count."}`);
+  } else if (typeof caps?.maxInputImages === "number") {
     lines.push(`Input images: ${caps.maxInputImages > 0 ? `up to ${caps.maxInputImages}` : "none — text-to-image only"}`);
   }
   if (caps?.resolutions?.length) {
@@ -146,7 +152,10 @@ async function chat({ prompt, model, system, temperature, max_tokens }) {
 }
 
 async function generateImage({ prompt, model, ...rest }) {
-  if (!prompt) throw new Error("`prompt` is required");
+  // A layer-decomposition model splits `image_url` and takes the prompt as an
+  // optional caption; every other image model needs the prompt. The gateway
+  // knows which is which and says so.
+  if (!prompt && !rest.image_url) throw new Error("`prompt` is required (or `image_url` alone, on a layer-decomposition model)");
   const res = await fetch(`${BASE}/api/v1/images/generations`, {
     method: "POST",
     headers: { ...headers(), "idempotency-key": crypto.randomUUID() },
@@ -156,12 +165,16 @@ async function generateImage({ prompt, model, ...rest }) {
     // live on the gateway and unreachable from any MCP client, silently. The
     // gateway validates against the model's capabilities and says what it
     // does not accept; this layer has no business having an opinion.
-    body: JSON.stringify({ model: model || "gpt-image-2", prompt, ...rest }),
+    body: JSON.stringify({ model: model || "gpt-image-2", ...(prompt ? { prompt } : {}), ...rest }),
   });
   if (!res.ok) throw new Error(await gwError(res));
   const json = await res.json();
-  const urls = (json.data ?? []).map((d) => d.url).filter(Boolean);
-  return urls.length ? urls.join("\n") : "(no image returned)";
+  // A layer stack prints one line per layer with what the model called it, so
+  // the agent can pick "headline" out of eight URLs without opening them.
+  const lines = (json.data ?? []).filter((d) => d.url).map((d) =>
+    d.layer ? `layer ${d.layer.index}${d.layer.name ? ` (${d.layer.name})` : ""}: ${d.url}` : d.url,
+  );
+  return lines.length ? lines.join("\n") : "(no image returned)";
 }
 
 async function generateVideo({ prompt, model, ...rest }) {
@@ -243,11 +256,11 @@ const TOOLS = [
   },
   {
     name: "ottoport_generate_image",
-    description: "Generate an image through an OttoPort image model (GPT Image, Nano Banana, …). Text-to-image, editing from one reference, or several references at once. Which a model accepts is in its capabilities — call ottoport_list_models. Returns image URL(s).",
+    description: "Generate an image through an OttoPort image model (GPT Image, Nano Banana, …). Text-to-image, editing from one reference, or several references at once. Also layer decomposition: qwen-image-layered and seedream-5.0-layers take `image_url` alone and return the image split into RGBA layers, one URL per layer, billed per layer. Which a model accepts is in its capabilities — call ottoport_list_models. Returns image URL(s).",
     inputSchema: {
       type: "object",
       properties: {
-        prompt: { type: "string" },
+        prompt: { type: "string", description: "What to generate. On a layer-decomposition model, an optional caption of the input image." },
         model: { type: "string", description: "Model id, e.g. gpt-image-2, nano-banana-2, nano-banana-pro. Default gpt-image-2." },
         size: { type: "string", description: 'Exact pixels, e.g. "1024x1024". Wins over `resolution`.' },
         // No `enum`: the vocabulary is "1k"/"2k"/"4k", but which of them a given
@@ -259,10 +272,10 @@ const TOOLS = [
         aspect_ratio: { type: "string", description: 'e.g. "16:9". Read only when `size` is absent.' },
         n: { type: "number" },
         seed: { type: "number" },
-        image_url: { type: "string", description: "One reference image, for editing and image-to-image." },
+        image_url: { type: "string", description: "One reference image, for editing and image-to-image — or, on a layer-decomposition model, the image to split." },
         reference_image_urls: { type: "array", items: { type: "string" }, description: "Several references — a character sheet, a product plus a scene — up to the model's maxInputImages. Call ottoport_list_models to see it." },
+        num_layers: { type: "number", description: "Layer models that let you choose (qwen-image-layered, 2–10): how many RGBA layers to split into. Each bills separately. seedream-5.0-layers decides for itself and refuses this." },
       },
-      required: ["prompt"],
     },
     handler: generateImage,
   },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ottoport",
-  "version": "1.4.1",
+  "version": "1.5.0",
   "description": "Claude Code plugin, CLI and MCP server for OttoPort — one OpenAI-compatible API for every LLM, image, video, and speech model.",
   "homepage": "https://ottoport.ai",
   "license": "MIT",

package/skills/ottoport/SKILL.md

@@ -16,7 +16,7 @@ and do not call the HTTP API, when a tool covers the job:
 | --- | --- |
 | `ottoport_list_models` | The catalog: what each model is for, and its rate in credits. Filter with `modality`, or pass `model` for ONE model in full — its modes, its resolution tiers with prices, and every parameter it reads. |
 | `ottoport_chat` | `prompt`, plus optional `model`, `system`, `temperature`, `max_tokens`. |
-| `ottoport_generate_image` | `prompt`, plus optional `model`, `size` or `resolution`, `aspect_ratio`, `n`, `seed`, `image_url` (one reference), `reference_image_urls` (several). |
+| `ottoport_generate_image` | `prompt`, plus optional `model`, `size` or `resolution`, `aspect_ratio`, `n`, `seed`, `image_url` (one reference), `reference_image_urls` (several). Layer decomposition too: `model` `qwen-image-layered` or `seedream-5.0-layers` with `image_url` alone (prompt optional, `num_layers` on Qwen) returns one URL per RGBA layer. |
 | `ottoport_generate_video` | `prompt`, plus optional `model`, `duration`, `resolution`, `aspect_ratio`, `seed`, `image_url` (first frame), `last_frame_url`, `reference_image_urls`. |
 | `ottoport_generate_speech` | `prompt` (the text), plus optional `model`, `voice`, `format`. |
 | `ottoport_generate_music` | `prompt`, plus optional `model`, `duration`, `format`. |
@@ -40,6 +40,13 @@ prints each one's menu beside its price:
 Images are the same idea with one axis: `image_url` for a single edit, or
 `reference_image_urls` for several at once, up to that model's limit.
 
+Two image models do not generate at all — they take an image apart.
+`qwen-image-layered` and `seedream-5.0-layers` split `image_url` into RGBA
+layers and return one URL per layer, bottom of the stack first; the prompt is
+an optional caption of the image, `num_layers` (2–10) is honoured by Qwen only,
+and **each layer returned bills separately**. Reach for them when the user wants
+a poster, ad or product shot as editable parts rather than a new picture.
+
 `resolution` is separate from all of it, and **it is priced** — a higher tier
 costs more. The vocabulary is `"480p"`/`"720p"`/`"1080p"`/`"2k"`/`"4k"` for
 video and `"1k"`/`"2k"`/`"4k"` for images, but **no model offers all of it**: