instantclips-mcp 1.4.0 → 1.5.0

package/README.es.md

@@ -94,7 +94,10 @@ Aplicación de Claude: añade un conector personalizado con esta dirección e in
 lo pida. ChatGPT en la web: activa el modo de desarrollador en Ajustes, Apps, Avanzado y añade la
 dirección como conector; en un espacio de trabajo Business o Enterprise, un administrador la
 publica como aplicación para todo el equipo. La aplicación de escritorio de ChatGPT acepta la misma
-dirección en Ajustes, Servidores MCP, y la comparte con Codex.
+dirección en Ajustes, Servidores MCP, y la comparte con Codex. En ChatGPT, adjunta las fotos del
+producto a la conversación y pide el vídeo: el plugin recibe los adjuntos directamente
+(`image_files` en `create_product_from_images`, `add_image_files` en `update_product`), así que
+allí las fotos no necesitan adaptador ni token.
 
 ### Cualquier otro cliente o agente MCP
 
@@ -168,7 +171,9 @@ acepta `image_paths` (rutas a los archivos, hasta 9, de 8 MB cada uno) en lugar
 `update_product` acepta `add_image_paths`. Los archivos se empaquetan y se envían directamente a
 InstantClips como fotos del producto; no se descarga nada de ningún sitio. Arrastra los archivos a
 un cliente que le pase sus rutas al asistente (Claude Code, Cursor, VS Code, agentes de terminal) y
-di qué quieres crear.
+di qué quieres crear. Envía `image_paths` solo: si en la misma llamada van también `image_urls`, el
+adaptador la rechaza en lugar de descartar las URL; añade las fotos alojadas después con
+`update_product`.
 
 ## Instrucciones para empezar
 

package/README.md

@@ -88,6 +88,9 @@ Claude app: add a custom connector with this address and sign in when it asks. C
 web: turn on Developer mode under Settings, Apps, Advanced, then add the address as a connector;
 on a Business or Enterprise workspace an admin publishes it as an app for everyone instead. The
 ChatGPT desktop app takes the same address under Settings, MCP servers, and shares it with Codex.
+In ChatGPT, attach the product photos to the conversation and ask for the video: the plugin takes
+attachments directly (`image_files` on `create_product_from_images`, `add_image_files` on
+`update_product`), so photos need no adapter and no token there.
 
 ### Any other MCP client or agent
 
@@ -158,6 +161,8 @@ that never left your laptop. Through the adapter, `create_product_from_images` t
 `add_image_paths`. The files are packaged and posted straight to InstantClips as the product's
 photos — nothing is downloaded from anywhere. Drag the files into a client that hands the
 assistant their paths (Claude Code, Cursor, VS Code, terminal agents) and say what you want made.
+Send `image_paths` on its own: with hosted `image_urls` in the same call the adapter refuses
+rather than dropping the URLs; add hosted photos afterwards with `update_product`.
 
 ## Starter prompts
 

package/README.zh-CN.md

@@ -78,7 +78,9 @@ url = "https://app.instantclips.ai/mcp"
 
 Claude 应用：用这个地址添加自定义连接器，按提示登录。网页版 ChatGPT：在 设置 › Apps › 高级 中
 开启开发者模式，再把地址添加为连接器；Business 或 Enterprise 工作区则由管理员发布为全员可用的
-应用。ChatGPT 桌面版在 设置 › MCP 服务器 中填入同一地址，并与 Codex 共享配置。
+应用。ChatGPT 桌面版在 设置 › MCP 服务器 中填入同一地址，并与 Codex 共享配置。在 ChatGPT 里，把商品照片
+作为附件添加到对话中并提出需求即可：插件会直接接收附件（`create_product_from_images` 的 `image_files`、
+`update_product` 的 `add_image_files`），因此在那里传照片不需要适配器，也不需要令牌。
 
 ### 其他 MCP 客户端或智能体
 
@@ -143,7 +145,8 @@ INSTANTCLIPS_TOKEN="your-token" npx -y instantclips-mcp --check --json
 `create_product_from_images` 可用 `image_paths`（文件路径，最多 9 个，每个不超过 8 MB）代替
 `image_urls`，`update_product` 可用 `add_image_paths`。文件会被打包并直接发送到 InstantClips 作为商品照片，
 不会从任何地方下载。把文件拖进会把路径交给助手的客户端（Claude Code、Cursor、VS Code、终端智能体），
-然后说出你想要的视频即可。
+然后说出你想要的视频即可。`image_paths` 请单独发送：同一次调用里若还带有 `image_urls`，适配器会拒绝，
+而不是悄悄丢弃这些 URL；托管照片可以之后用 `update_product` 添加。
 
 ## 入门提示语
 

package/bin/instantclips-mcp.js

@@ -141,6 +141,14 @@ function writeDiagnostic(failure) {
 // serves. The files are read and posted as multipart to the hosted upload
 // endpoints under the same bearer — packaged and uploaded, never downloaded
 // from anywhere.
+//
+// The hosted server also takes the conversation's attachments — `image_files`
+// and `add_image_files`, declared for ChatGPT through `_meta["openai/fileParams"]`
+// — which a stdio client has no way to fill. Those two parameters and the
+// `_meta` are removed from the schemas served here, so the model sees one
+// local-file parameter and not a hosted one it cannot use. A call that sends
+// `image_paths` together with hosted URLs is refused rather than uploading
+// the files and dropping the URLs, which is what 1.4.0 did.
 const MAX_LOCAL_IMAGES = 9;
 const MAX_LOCAL_IMAGE_BYTES = 8 * 1024 * 1024;
 const LOCAL_IMAGE_TYPES = {
@@ -159,14 +167,17 @@ const LOCAL_UPLOAD_PARAMETERS = {
   create_product_from_images: {
     name: "image_paths",
     replaces: "image_urls",
+    hides: "image_files",
     description:
       `Paths to photos on this machine, up to ${MAX_LOCAL_IMAGES}, ${MAX_LOCAL_IMAGE_BYTES / 1024 / 1024} MB each. ` +
-      "This adapter uploads the files itself; use it instead of image_urls for local photos.",
+      "This adapter uploads the files itself; use it instead of image_urls for local photos, never with them.",
     note:
-      "Through this adapter, `image_paths` (files on this machine) can replace `image_urls`; the files are uploaded directly.",
+      "Through this adapter, `image_paths` (files on this machine) replaces `image_urls`; the files are uploaded directly. " +
+      "Send one or the other, not both.",
   },
   update_product: {
     name: "add_image_paths",
+    hides: "add_image_files",
     description:
       `Paths to photos on this machine to add, up to ${MAX_LOCAL_IMAGES}, ${MAX_LOCAL_IMAGE_BYTES / 1024 / 1024} MB each. ` +
       "This adapter uploads the files itself. Send it with product_id alone; other fields go in a separate call.",
@@ -180,18 +191,35 @@ function withLocalUploads(tools) {
     const extra = LOCAL_UPLOAD_PARAMETERS[tool.name];
     if (!extra) return tool;
     const schema = tool.inputSchema || { type: "object", properties: {} };
+    const { [extra.hides]: _hidden, ...kept } = schema.properties || {};
     const properties = {
-      ...schema.properties,
-      [extra.name]: { type: "array", items: { type: "string" }, description: extra.description },
+      ...kept,
+      [extra.name]: { type: "array", items: { type: "string" }, minItems: 1, description: extra.description },
     };
     const inputSchema = { ...schema, properties };
     if (extra.replaces && Array.isArray(schema.required)) {
       inputSchema.required = schema.required.filter((key) => key !== extra.replaces);
     }
-    return { ...tool, description: `${tool.description}\n\n${extra.note}`, inputSchema };
+    const { _meta, ...rest } = tool;
+    return {
+      ...rest,
+      ...withoutFileParam(_meta, extra.hides),
+      description: `${tool.description}\n\n${extra.note}`,
+      inputSchema,
+    };
   });
 }
 
+// → `{ _meta }` with the hidden file parameter's declaration removed, or `{}`
+// when nothing else was in it.
+function withoutFileParam(meta, hidden) {
+  if (!meta) return {};
+  const { "openai/fileParams": fileParams, ...others } = meta;
+  const remaining = (fileParams || []).filter((name) => name !== hidden);
+  const kept = remaining.length ? { ...others, "openai/fileParams": remaining } : others;
+  return Object.keys(kept).length ? { _meta: kept } : {};
+}
+
 function hasPaths(value) {
   return Array.isArray(value) && value.length > 0;
 }
@@ -250,13 +278,31 @@ async function uploadForm(config, suffix, form) {
     error.data = { status: 401 };
     throw error;
   }
-  return { content: [{ type: "text", text }], isError: !response.ok };
+  if (!response.ok) return { content: [{ type: "text", text }], isError: true };
+  // The hosted tools declare output schemas, and a validating client refuses
+  // a result that has one but no structuredContent — so the upload's payload
+  // (the same product object get_product returns) travels both ways.
+  const result = { content: [{ type: "text", text }], isError: false };
+  try {
+    const parsed = JSON.parse(text);
+    if (parsed && typeof parsed === "object" && !Array.isArray(parsed)) result.structuredContent = parsed;
+  } catch {
+    // Not JSON: the text block stands alone, as it always did.
+  }
+  return result;
 }
 
 // → a tool result when the call is one this adapter answers itself, else null.
 async function callLocally(config, params) {
   const args = params.arguments ?? {};
   if (params.name === "create_product_from_images" && hasPaths(args.image_paths)) {
+    const hosted = ["image_urls", "image_files"].filter((key) => Array.isArray(args[key]) && args[key].length > 0);
+    if (hosted.length > 0) {
+      throw uploadError(
+        `Send image_paths alone: ${hosted.join(", ")} would be dropped. Create the product from image_paths, ` +
+          "then add hosted photos with update_product's add_image_urls.",
+      );
+    }
     const form = new FormData();
     for (const key of ["name", "description", "creator_note", "brand_id"]) {
       if (args[key] !== undefined && args[key] !== null && args[key] !== "") form.append(key, String(args[key]));