ottoport 1.6.0 → 1.8.0

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": "1.6.0",
+  "version": "1.8.0",
   "author": {
     "name": "LITBOX LLC",
     "email": "support@ottoport.ai"

package/README.md

@@ -20,7 +20,7 @@ ottoport models [--modality chat|image|video|tts|music] [--json]
 ottoport chat  "<prompt>" [--model claude-sonnet-5] [--system "..."] [--no-stream]
                           [--temperature 0.7] [--max-tokens 512]
 ottoport image "<prompt>" [--model gpt-image-2] [--size 1024x1024 | --resolution 2k]
-                          [--aspect-ratio 16:9] [--n 1] [--seed 7]
+                          [--quality high] [--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]
@@ -90,6 +90,11 @@ 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.
 
+The GPT Image family reads a `quality` on top of that — `low`, `medium` (the
+default every catalog rate is measured at), `high`, and on `gpt-image-2.5` and
+`gpt-image-2.5-flare` also `xhigh` and `max`. It is how much compute the model
+spends, and the ladder is steep: `max` bills sixteen times `medium`.
+
 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
@@ -133,8 +138,8 @@ https://ottoport.ai/docs#webhooks.
 - Video generation takes a minute or more and the call blocks until the job is
   terminal, unless a webhook is given or `--no-wait` is passed. A retry is a
   second billable generation, not a resumption.
-- `resolution` is a price multiplier, not a formatting hint. Leave it unset to
-  bill at the model's base tier.
+- `resolution` and `quality` are price multipliers, not formatting hints. Leave
+  them unset to bill at the model's base tier.
 - Requests are billed against the prepaid balance on your OttoPort account.
 
 Docs: <https://ottoport.ai/docs> · Support: support@ottoport.ai

package/cli/ottoport.mjs

@@ -5,6 +5,7 @@
 //
 //   ottoport models [--modality chat|image|video]
 //   ottoport chat   "prompt" [--model claude-sonnet-5] [--system "..."] [--no-stream]
+//                   [--video <url|file>]   ask about a video (defaults to video-understanding)
 //   ottoport image  "prompt" [--model gpt-image-2] [--resolution 2k] [--ref a.png,b.png] [--out img.png]
 //   ottoport video  "prompt" [--model kling-3.0] [--duration 5] [--out clip.mp4]
 //   ottoport speech "text" [--model gpt-4o-mini-tts] [--voice alloy] [--out speech.mp3]
@@ -17,7 +18,7 @@
 //   The public CLI always targets https://ottoport.ai unless --url is passed
 //   explicitly for local development or a self-hosted gateway.
 
-import { writeFile } from "node:fs/promises";
+import { readFile, stat, writeFile } from "node:fs/promises";
 
 const BASE = "https://ottoport.ai";
 
@@ -152,11 +153,13 @@ async function cmdChat(prompt, flags) {
   const { baseUrl, apiKey } = config(flags);
   const messages = [];
   if (flags.system) messages.push({ role: "system", content: flags.system });
-  messages.push({ role: "user", content: prompt });
+  const videos = await videoParts(flags);
+  messages.push({ role: "user", content: videos.length ? [{ type: "text", text: prompt }, ...videos] : prompt });
 
   const stream = flags.stream !== false;
   const body = {
-    model: last(flags.model) || "claude-sonnet-5",
+    // A video question goes to the model built for it unless one is named.
+    model: last(flags.model) || (videos.length ? "video-understanding" : "claude-sonnet-5"),
     messages,
     stream,
     ...(flags.temperature ? { temperature: Number(last(flags.temperature)) } : {}),
@@ -202,6 +205,39 @@ async function cmdChat(prompt, flags) {
   process.stdout.write("\n");
 }
 
+const VIDEO_TYPES = { ".mp4": "video/mp4", ".m4v": "video/mp4", ".mov": "video/quicktime", ".webm": "video/webm", ".mpeg": "video/mpeg", ".mpg": "video/mpeg" };
+
+/**
+ * `--video`, once or several times (or comma-separated), as chat content parts.
+ *
+ * A URL is sent as-is: the gateway's transport fetches it, so it must be
+ * publicly reachable. A local file is inlined as a base64 data URL, which is
+ * convenient for trying a clip but costs a third more than the file on the
+ * wire — for anything large, host it and pass the URL.
+ */
+async function videoParts(flags) {
+  const list = (Array.isArray(flags.video) ? flags.video : [flags.video])
+    .filter((value) => typeof value === "string")
+    .flatMap((value) => value.split(",").map((one) => one.trim()))
+    .filter(Boolean);
+  const parts = [];
+  for (const source of list) {
+    if (/^(https?:|data:)/i.test(source)) {
+      parts.push({ type: "video_url", video_url: { url: source } });
+      continue;
+    }
+    const info = await stat(source).catch(() => null);
+    if (!info?.isFile()) die(`--video ${source}: not a URL and no such file`);
+    const ext = source.slice(source.lastIndexOf(".")).toLowerCase();
+    const type = VIDEO_TYPES[ext];
+    if (!type) die(`--video ${source}: unsupported type ${ext || "(none)"}; use ${Object.keys(VIDEO_TYPES).join(", ")}`);
+    const mb = info.size / 1e6;
+    if (mb > 20) console.error(`ottoport: ${source} is ${mb.toFixed(0)}MB; inlined it will be ~${(mb * 4 / 3).toFixed(0)}MB and may be refused — host it and pass the URL instead`);
+    parts.push({ type: "video_url", video_url: { url: `data:${type};base64,${(await readFile(source)).toString("base64")}` } });
+  }
+  return parts;
+}
+
 async function download(url, out) {
   const res = await fetch(url);
   if (!res.ok) die(`could not download ${url}: ${res.status}`);
@@ -236,6 +272,7 @@ async function cmdImage(prompt, flags) {
     ...(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.quality) ? { quality: last(flags.quality) } : {}),
     ...(last(flags["aspect-ratio"]) ? { aspect_ratio: flags["aspect-ratio"] } : {}),
     ...(last(flags.n) ? { n: Number(last(flags.n)) } : {}),
     ...(last(flags.seed) ? { seed: Number(last(flags.seed)) } : {}),
@@ -376,8 +413,9 @@ Usage:
   ottoport models [--modality chat|image|video|tts|music] [--json]
   ottoport models <model-id> [--json]     what that one model accepts
   ottoport chat  "<prompt>" [--model claude-sonnet-5] [--system "..."] [--no-stream]
+                            [--video <url|file>[,<url|file>…]]   ask about a video
                             [--temperature 0.7] [--max-tokens 512]
-  ottoport image "<prompt>" [--model gpt-image-2] [--size 1024x1024 | --resolution 2k]
+  ottoport image "<prompt>" [--model gpt-image-2] [--size 1024x1024 | --resolution 2k] [--quality high]
                             [--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]
@@ -408,6 +446,7 @@ Examples:
   ottoport models --modality image      # prints what each one accepts
   ottoport models veo-3.1               # its modes, tiers, and every parameter
   ottoport chat "explain MCP in one line" --model claude-haiku-4.5
+  ottoport chat "break this clip down shot by shot" --video https://example.com/clip.mp4
   ottoport image "isometric city at dusk" --model gpt-image-2 --out city.png
   ottoport video "cinematic product reveal" --model seedance-2.0-fast --resolution 1080p
   ottoport video "the logo unfolds" --image start.png --last-frame end.png

package/commands/image.md

@@ -1,6 +1,6 @@
 ---
 description: Generate an image through the OttoPort gateway
-argument-hint: "<prompt> [--model <id>] [--resolution 2k] [--ref <url>…]"
+argument-hint: "<prompt> [--model <id>] [--resolution 2k] [--quality high] [--ref <url>…]"
 allowed-tools: mcp__ottoport__ottoport_generate_image, mcp__ottoport__ottoport_list_models, Bash(curl:*)
 ---
 
@@ -15,6 +15,9 @@ Request: $ARGUMENTS
 - Detail is `resolution` (`1k`/`2k`/`4k`), optionally with `aspect_ratio`; pass
   `size` instead when the user named exact pixels. A higher tier costs more, so
   send one only when the request asks for it.
+- On the GPT Image models, `quality` buys compute rather than pixels — `low`,
+  `medium` (the default), `high`, and `xhigh`/`max` on `gpt-image-2.5`. It is
+  billed steeply, so send it only when the user asked for a draft or a finish.
 - Models differ in how many references and which resolutions they take. If a
   request needs more than one reference, or 4k, check `ottoport_list_models`
   first — it prints each model's menu — rather than spending a failed call.

package/mcp/server.mjs

@@ -290,6 +290,7 @@ const TOOLS = [
         // at best. The model's own menu is in `ottoport_list_models`.
         resolution: { type: "string", description: 'Output detail — "1k", "2k" or "4k", but only the tiers this model lists in ottoport_list_models. Priced accordingly. Pair with `aspect_ratio` when you know the shape but not the pixel vocabulary; an explicit `size` wins.' },
         aspect_ratio: { type: "string", description: 'e.g. "16:9". Read only when `size` is absent.' },
+        quality: { type: "string", description: 'How much compute the model spends, on the GPT Image family only — "low", "medium" (the default), "high", and on gpt-image-2.5 "xhigh" and "max". Priced by tier, and the ladder is steep: `max` costs sixteen times `medium`. Check the model\'s `qualities` in ottoport_list_models.' },
         n: { type: "number" },
         seed: { type: "number" },
         image_url: { type: "string", description: "One reference image, for editing and image-to-image — or, on a layer-decomposition model, the image to split." },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "ottoport",
-  "version": "1.6.0",
+  "version": "1.8.0",
   "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",
   "repository": {

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 and USD. 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). 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_image` | `prompt`, plus optional `model`, `size` or `resolution`, `quality` (GPT Image only), `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`, `webhook_url` (returns the job id at once and POSTs the result there; the secret comes from `webhook_secret` or the server's `OTTOPORT_WEBHOOK_SECRET`). |
 | `ottoport_video_webhook` | `job_id`, plus `action` `status` (every delivery attempt and the endpoint's response) or `retry` (redeliver a finished job's webhook now). |
 | `ottoport_generate_speech` | `prompt` (the text), plus optional `model`, `voice`, `format`. |
@@ -55,6 +55,12 @@ video and `"1k"`/`"2k"`/`"4k"` for images, but **no model offers all of it**:
 `gpt-image-1.5` only 1k. Same for how many references a model takes —
 `nano-banana-lite` accepts 3, `gpt-image-2` accepts 16.
 
+`quality` is a second priced dial, on the GPT Image family alone: `low`,
+`medium` (the default, and the tier every catalog rate was measured at),
+`high`, and `xhigh`/`max` on `gpt-image-2.5` and `gpt-image-2.5-flare`. It buys
+compute, not pixels, and the ladder is steep — `max` bills sixteen times
+`medium`. A model's tiers are its `qualities` in `ottoport_list_models`.
+
 So read the model's own menu before naming a mode, a resolution or a second
 reference image: `ottoport_list_models` with `model` set to the one you are
 about to call answers with exactly that — its modes, what each resolution tier