@kolbo/mcp 1.113.1 → 1.113.3

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@kolbo/mcp",
-  "version": "1.113.1",
+  "version": "1.113.3",
   "description": "Kolbo AI MCP Server - Generate images, videos, music, speech, and sound effects from Claude Code",
   "main": "src/index.js",
   "bin": {

package/src/apps/index.js

@@ -366,6 +366,9 @@ async function modelInfoMap(client) {
 const DISPLAY_ALIASES = {
   'flux-3-draft-enhance': { base: 'seedance-2-5', name: 'Seedance 2.5 Draft Enhance' },
   'video-draft-enhance': { base: 'seedance-2-5', name: 'Seedance 2.5 Draft Enhance' },
+  // Billing SKU every ElevenLabs CLONED voice reports; its version shows in the
+  // separate "Model V4/V3/V2" chip, so the model chip is just the provider.
+  'fal_elevenlabs_tts': { base: 'eleven_v4', name: 'ElevenLabs' },
 };
 // Operations whose engine is picked server-side when no model is passed.
 const OPERATION_ENGINES = { draft_enhance: 'flux-3-draft-enhance' };
@@ -530,10 +533,16 @@ async function resolveCatalogAspectRatio(client, modelId, requested, type) {
 
 // Same shape and TTL as the model catalog above: the voice catalog is stable,
 // and a per-generation /v1/voices round trip would be paid on every speech card.
-const voiceCache = new Map(); // apiBase → { at, byKey }
+// Keyed per caller, not per server: /v1/voices includes the caller's own cloned
+// voices, so one shared entry (the remote connector serves every user from one
+// process) left everyone else's clones unresolved for the whole TTL.
+const voiceCache = new Map(); // apiBase + key hash → { at, byKey }
 
 async function voiceInfoMap(client) {
-  const cacheKey = client.apiBase || 'default';
+  const keyHash = client.apiKey
+    ? crypto.createHash('sha256').update(String(client.apiKey)).digest('hex').slice(0, 16)
+    : 'anon';
+  const cacheKey = `${client.apiBase || 'default'}|${keyHash}`;
   const hit = voiceCache.get(cacheKey);
   if (hit && Date.now() - hit.at < ICON_TTL_MS) return hit.byKey;
   const byKey = new Map();

package/src/tools/generate.js

@@ -1201,7 +1201,8 @@ function registerGenerateTools(server, client, options = {}) {
       // the card was labelling those "Smart Select", which is not even a text-to-
       // speech option, and naming the voice the caller typed rather than the one
       // that spoke. Same resolution addDisplayNames does for the polling path.
-      const ranVoice = (await voiceInfo(client, result.result?.voice).catch(() => null)) || voiceRecord;
+      const ranVoice = statusVoice(result.result)
+        || (await voiceInfo(client, result.result?.voice).catch(() => null)) || voiceRecord;
       return uiCompleted({
         tool: 'generate_speech', kind: 'audio', gen, client, prompt: text,
         model: result.result?.model || model,
@@ -1314,13 +1315,21 @@ function registerGenerateTools(server, client, options = {}) {
       r.model_name = (info && info.name) || r.model;
       r.model_icon = (info && info.icon) || null;
     }
-    if (r.voice) {
+    // The API names a cloned voice itself (voice_name/voice_thumbnail): clones
+    // are not reliably in the cached catalog, so keep the API's answer.
+    if (r.voice && !statusVoice(r)) {
       const v = await voiceInfo(client, r.voice).catch(() => null);
       r.voice_name = (v && v.name) || r.voice;
       r.voice_thumbnail = (v && v.thumbnail) || null;
     }
   }
 
+  function statusVoice(r) {
+    if (!r || !r.voice_name) return null;
+    const id = /^custom_/i.test(String(r.requested_voice_id || '')) ? r.requested_voice_id : r.voice;
+    return { id, name: r.voice_name, thumbnail: r.voice_thumbnail || null };
+  }
+
   // ─── get_generation_status ─────────────────────────────────
   server.tool(
     'get_generation_status',
@@ -1941,6 +1950,18 @@ function registerGenerateTools(server, client, options = {}) {
       if (!source) throw new Error('source is required (URL or absolute local path to image/video)');
       if (!audio) throw new Error('audio is required (URL or absolute local path to audio file)');
 
+      // Upload local files first so the card can show the face source and the
+      // audio from the same CDN URLs the API receives (multipart left the card
+      // with no references at all).
+      ({ source, audio } = await rehostLocalPaths(client, { source, audio, project_id }, { allowLocalFiles: !options.remote }));
+      const isVideoUrl = (u) => /\.(mp4|mov|webm|mkv|avi|m4v)(\?|$)/i.test(u);
+      const httpRef = (u) => (typeof u === 'string' && /^https?:\/\//i.test(u) ? [u] : []);
+      const lipsyncRefs = {
+        reference_images: httpRef(source).filter((u) => !isVideoUrl(u)),
+        reference_videos: httpRef(source).filter(isVideoUrl),
+        reference_audio: httpRef(audio),
+      };
+
       const sourceIsUrl = typeof source === 'string' && /^https?:\/\//i.test(source);
       const audioIsUrl = typeof audio === 'string' && /^https?:\/\//i.test(audio);
 
@@ -1996,8 +2017,7 @@ function registerGenerateTools(server, client, options = {}) {
 
       if (returnsImmediately()) return submittedResult({
         tool: 'generate_lipsync', kind: 'video', gen: startResponse, client, model,
-        prompt: text_prompt, settings: { mode: 'lipsync' },
-        reference_images: sourceIsUrl && !/\.(mp4|mov|webm|mkv|avi|m4v)(\?|$)/i.test(source) ? [source] : [],
+        prompt: text_prompt, settings: { mode: 'lipsync' }, ...lipsyncRefs,
       });
 
       const poll = await pollOrTimedOut(client, startResponse.generation_id, {
@@ -2009,9 +2029,7 @@ function registerGenerateTools(server, client, options = {}) {
 
       return uiCompleted({
         tool: 'generate_lipsync', kind: 'video', gen: startResponse, client, model,
-        prompt: text_prompt, settings: { mode: 'lipsync' },
-        reference_images: sourceIsUrl && !/\.(mp4|mov|webm|mkv|avi|m4v)(\?|$)/i.test(source) ? [source] : [],
-        reference_videos: sourceIsUrl && /\.(mp4|mov|webm|mkv|avi|m4v)(\?|$)/i.test(source) ? [source] : [],
+        prompt: text_prompt, settings: { mode: 'lipsync' }, ...lipsyncRefs,
         urls: result.result?.urls || [],
         thumbnail_url: result.result?.thumbnail_url || null,
         duration: result.result?.duration || null,