feat: add OpenRouter + direct OpenAI as alternative LLM providers

lib/agents/CompactAgent/createCompactAgent.ts

@@ -1,5 +1,6 @@
 import { ToolLoopAgent, stepCountIs } from "ai";
 import { LIGHTWEIGHT_MODEL } from "@/lib/const";
+import { createModel } from "@/lib/ai/createModel";
 
 const DEFAULT_INSTRUCTIONS = `You are a conversation summarizer. Create a concise summary of the conversation that:
 - Preserves key information, decisions, and action items
@@ -17,7 +18,7 @@ Respond with only the summary text, no additional commentary.`;
  */
 export function createCompactAgent(customInstructions?: string) {
   return new ToolLoopAgent({
-    model: LIGHTWEIGHT_MODEL,
+    model: createModel(LIGHTWEIGHT_MODEL),
     instructions: customInstructions || DEFAULT_INSTRUCTIONS,
     stopWhen: stepCountIs(1),
   });

lib/agents/EmailReplyAgent/createEmailReplyAgent.ts

@@ -1,6 +1,7 @@
 import { Output, ToolLoopAgent, stepCountIs } from "ai";
 import { z } from "zod";
 import { LIGHTWEIGHT_MODEL, INBOUND_EMAIL_DOMAIN } from "@/lib/const";
+import { createModel } from "@/lib/ai/createModel";
 
 const replyDecisionSchema = z.object({
   shouldReply: z.boolean().describe("Whether the Recoup AI assistant should reply to this email"),
@@ -22,7 +23,7 @@ Rules (check in this order):
  */
 export function createEmailReplyAgent() {
   return new ToolLoopAgent({
-    model: LIGHTWEIGHT_MODEL,
+    model: createModel(LIGHTWEIGHT_MODEL),
     instructions,
     output: Output.object({ schema: replyDecisionSchema }),
     stopWhen: stepCountIs(1),

lib/agents/content/createContentPromptAgent.ts

@@ -1,6 +1,7 @@
 import { Output, ToolLoopAgent, stepCountIs } from "ai";
 import { z } from "zod";
 import { LIGHTWEIGHT_MODEL } from "@/lib/const";
+import { createModel } from "@/lib/ai/createModel";
 import { CONTENT_TEMPLATES, DEFAULT_CONTENT_TEMPLATE } from "@/lib/content/contentTemplates";
 import { CAPTION_LENGTHS } from "@/lib/content/captionLengths";
 import { songsSchema } from "@/lib/content/songsSchema";
@@ -68,7 +69,7 @@ Defaults: lipsync=${DEFAULT_CONTENT_PROMPT_FLAGS.lipsync}, batch=${DEFAULT_CONTE
  */
 export function createContentPromptAgent() {
   return new ToolLoopAgent({
-    model: LIGHTWEIGHT_MODEL,
+    model: createModel(LIGHTWEIGHT_MODEL),
     instructions,
     output: Output.object({ schema: contentPromptFlagsSchema }),
     stopWhen: stepCountIs(1),

lib/agents/generalAgent/getGeneralAgent.ts

@@ -3,6 +3,7 @@ import { AnthropicProviderOptions } from "@ai-sdk/anthropic";
 import { GoogleGenerativeAIProviderOptions } from "@ai-sdk/google";
 import { OpenAIResponsesProviderOptions } from "@ai-sdk/openai";
 import { DEFAULT_MODEL } from "@/lib/const";
+import { createModel } from "@/lib/ai/createModel";
 import { RoutingDecision } from "@/lib/chat/types";
 import { extractImageUrlsFromMessages } from "@/lib/messages/extractImageUrlsFromMessages";
 import { buildSystemPromptWithImages } from "@/lib/chat/buildSystemPromptWithImages";
@@ -51,7 +52,7 @@ export default async function getGeneralAgent(body: ChatRequestBody): Promise<Ro
   const instructions = buildSystemPromptWithImages(baseSystemPrompt, imageUrls);
 
   const tools = await setupToolsForRequest(body);
-  const model = bodyModel || DEFAULT_MODEL;
+  const model = createModel(bodyModel || DEFAULT_MODEL);
   const stopWhen = stepCountIs(111);
 
   const agent = new ToolLoopAgent({

lib/ai/createModel.ts

@@ -0,0 +1,39 @@
+import type { LanguageModel } from "ai";
+import { gateway } from "@ai-sdk/gateway";
+import { createOpenRouter } from "@openrouter/ai-sdk-provider";
+import { openai } from "@ai-sdk/openai";
+
+/**
+ * Resolves a model string (e.g. "openai/gpt-5-mini") to a LanguageModel,
+ * routing through whichever provider is configured.
+ *
+ * Priority: Vercel AI Gateway > OpenRouter > direct OpenAI.
+ */
+export function createModel(modelId: string): LanguageModel {
+  // Vercel AI Gateway — existing production behavior
+  if (process.env.VERCEL_AI_GATEWAY_API_KEY || process.env.VERCEL_OIDC_TOKEN) {
+    return gateway(modelId);
+  }
+
+  // OpenRouter — supports "provider/model" format natively
+  if (process.env.OPENROUTER_API_KEY) {
+    const openrouter = createOpenRouter({ apiKey: process.env.OPENROUTER_API_KEY });
+    return openrouter(modelId);
+  }
+
+  // Direct OpenAI — only supports openai/* models
+  if (process.env.OPENAI_API_KEY) {
+    if (!modelId.startsWith("openai/") && modelId.includes("/")) {
+      throw new Error(
+        `Model "${modelId}" is not an OpenAI model. Direct OpenAI mode only supports openai/* models. ` +
+          `Use OPENROUTER_API_KEY or VERCEL_AI_GATEWAY_API_KEY for multi-provider support.`,
+      );
+    }
+    const bareModel = modelId.startsWith("openai/") ? modelId.slice(7) : modelId;
+    return openai(bareModel);
+  }
+
+  throw new Error(
+    "No LLM provider configured. Set VERCEL_AI_GATEWAY_API_KEY, OPENROUTER_API_KEY, or OPENAI_API_KEY.",
+  );
+}

lib/ai/generateArray.ts

@@ -1,5 +1,6 @@
 import { generateObject } from "ai";
 import { DEFAULT_MODEL } from "@/lib/const";
+import { createModel } from "./createModel";
 import { z } from "zod";
 
 export interface GenerateArrayResult {
@@ -15,7 +16,7 @@ const generateArray = async ({
   prompt: string;
 }): Promise<GenerateArrayResult[]> => {
   const result = await generateObject({
-    model: DEFAULT_MODEL,
+    model: createModel(DEFAULT_MODEL),
     system,
     prompt,
     output: "array",

lib/ai/generateText.ts

@@ -1,5 +1,6 @@
 import { generateText as generate } from "ai";
 import { DEFAULT_MODEL } from "@/lib/const";
+import { createModel } from "./createModel";
 
 const generateText = async ({
   system,
@@ -12,7 +13,7 @@ const generateText = async ({
 }) => {
   const result = await generate({
     system,
-    model: model || DEFAULT_MODEL,
+    model: createModel(model || DEFAULT_MODEL),
     prompt,
   });
 

lib/ai/getAvailableModels.ts

@@ -1,16 +1,45 @@
 import { gateway, GatewayLanguageModelEntry } from "@ai-sdk/gateway";
 import isEmbedModel from "./isEmbedModel";
+import { DEFAULT_MODEL, LIGHTWEIGHT_MODEL } from "@/lib/const";
 
 /**
- * Returns the list of available LLMs from the Vercel AI Gateway.
- * Filters out embed models that are not suitable for chat.
+ * Default model list for non-gateway providers (OpenRouter, direct OpenAI).
+ * Returned only when the Vercel AI Gateway is not configured.
+ */
+const DEFAULT_MODELS: GatewayLanguageModelEntry[] = [
+  {
+    id: DEFAULT_MODEL,
+    name: "GPT-5 Mini",
+    description: "Default model for chat and generation",
+    pricing: { input: "0.0001", output: "0.0004" },
+    specification: { specificationVersion: "v2", provider: "openai", modelId: DEFAULT_MODEL },
+  },
+  {
+    id: LIGHTWEIGHT_MODEL,
+    name: "GPT-4o Mini",
+    description: "Lightweight model for simple tasks",
+    pricing: { input: "0.00015", output: "0.0006" },
+    specification: { specificationVersion: "v2", provider: "openai", modelId: LIGHTWEIGHT_MODEL },
+  },
+];
+
+/**
+ * Returns the list of available LLMs.
+ * Uses Vercel AI Gateway when configured, otherwise returns a default list.
  */
 export const getAvailableModels = async (): Promise<GatewayLanguageModelEntry[]> => {
-  try {
-    const apiResponse = await gateway.getAvailableModels();
-    const gatewayModels = apiResponse.models.filter(m => !isEmbedModel(m));
-    return gatewayModels;
-  } catch {
-    return [];
+  // Use Vercel AI Gateway when configured
+  if (process.env.VERCEL_AI_GATEWAY_API_KEY) {
+    try {
+      const apiResponse = await gateway.getAvailableModels();
+      const gatewayModels = apiResponse.models.filter(m => !isEmbedModel(m));
+      return gatewayModels;
+    } catch (error) {
+      console.error("[getAvailableModels] Gateway fetch failed:", error);
+      return [];
+    }
   }
+
+  // Fallback for OpenRouter or direct OpenAI
+  return DEFAULT_MODELS;
 };

lib/catalog/analyzeCatalogBatch.ts

@@ -2,6 +2,7 @@ import { z } from "zod";
 import { generateObject } from "ai";
 import type { CatalogSongWithArtists } from "@/lib/supabase/catalog_songs/selectCatalogSongsWithArtists";
 import { DEFAULT_MODEL } from "@/lib/const";
+import { createModel } from "@/lib/ai/createModel";
 
 /**
  * Analyzes a single batch of catalog songs using AI to filter by criteria
@@ -17,7 +18,7 @@ export async function analyzeCatalogBatch(
 ): Promise<CatalogSongWithArtists[]> {
   // Use AI to select relevant songs from this batch
   const { object } = await generateObject({
-    model: DEFAULT_MODEL,
+    model: createModel(DEFAULT_MODEL),
     schema: z.object({
       selected_song_isrcs: z
         .array(z.string())

lib/chat/toolChains/toolChains.ts

@@ -1,6 +1,7 @@
 import { LanguageModel, ModelMessage } from "ai";
 import { createReleaseReportToolChain } from "./create_release_report/createReleaseReportToolChain";
 import { createNewArtistToolChain } from "./createNewArtistToolChain";
+import { createModel } from "@/lib/ai/createModel";
 
 export type ToolChainItem = {
   toolName: string;
@@ -18,22 +19,22 @@ export type PrepareStepResult = {
 // Forced toolChoice is incompatible with Anthropic extended thinking.
 // Every tool used in a chain must have a model here to avoid the conflict.
 export const TOOL_MODEL_MAP: Record<string, LanguageModel> = {
-  update_account_info: "gemini-2.5-pro",
-  get_spotify_search: "openai/gpt-5.4-mini",
-  update_artist_socials: "openai/gpt-5.4-mini",
-  artist_deep_research: "openai/gpt-5.4-mini",
-  spotify_deep_research: "openai/gpt-5.4-mini",
-  get_artist_socials: "openai/gpt-5.4-mini",
-  get_spotify_artist_top_tracks: "openai/gpt-5.4-mini",
-  get_spotify_artist_albums: "openai/gpt-5.4-mini",
-  get_spotify_album: "openai/gpt-5.4-mini",
-  search_web: "openai/gpt-5.4-mini",
-  generate_txt_file: "openai/gpt-5.4-mini",
-  create_segments: "openai/gpt-5.4-mini",
-  youtube_login: "openai/gpt-5.4-mini",
-  web_deep_research: "openai/gpt-5.4-mini",
-  create_knowledge_base: "openai/gpt-5.4-mini",
-  send_email: "openai/gpt-5.4-mini",
+  update_account_info: createModel("google/gemini-2.5-pro"),
+  get_spotify_search: createModel("openai/gpt-5.4-mini"),
+  update_artist_socials: createModel("openai/gpt-5.4-mini"),
+  artist_deep_research: createModel("openai/gpt-5.4-mini"),
+  spotify_deep_research: createModel("openai/gpt-5.4-mini"),
+  get_artist_socials: createModel("openai/gpt-5.4-mini"),
+  get_spotify_artist_top_tracks: createModel("openai/gpt-5.4-mini"),
+  get_spotify_artist_albums: createModel("openai/gpt-5.4-mini"),
+  get_spotify_album: createModel("openai/gpt-5.4-mini"),
+  search_web: createModel("openai/gpt-5.4-mini"),
+  generate_txt_file: createModel("openai/gpt-5.4-mini"),
+  create_segments: createModel("openai/gpt-5.4-mini"),
+  youtube_login: createModel("openai/gpt-5.4-mini"),
+  web_deep_research: createModel("openai/gpt-5.4-mini"),
+  create_knowledge_base: createModel("openai/gpt-5.4-mini"),
+  send_email: createModel("openai/gpt-5.4-mini"),
 };
 
 // Map trigger tool -> sequence AFTER trigger

lib/evals/scorers/CatalogAvailability.ts

@@ -1,5 +1,6 @@
 import { DEFAULT_MODEL } from "@/lib/consts";
 import { generateObject } from "ai";
+import { createModel } from "@/lib/ai/createModel";
 import { getCatalogDataAsCSV } from "@/lib/catalog/getCatalogDataAsCSV";
 import { z } from "zod";
 
@@ -18,7 +19,7 @@ export const CatalogAvailability = async ({
     const catalog = await getCatalogDataAsCSV(catalogId);
 
     const result = await generateObject({
-      model: DEFAULT_MODEL,
+      model: createModel(DEFAULT_MODEL),
       system: `You are a music catalog analyst. Your job is to analyze song recommendations and determine which ones are available in the provided music catalog.
 
 Instructions:

lib/evals/scorers/QuestionAnswered.ts

@@ -1,5 +1,6 @@
 import { DEFAULT_MODEL } from "@/lib/consts";
 import { generateObject } from "ai";
+import { createModel } from "@/lib/ai/createModel";
 import { z } from "zod";
 
 /**
@@ -17,7 +18,7 @@ export const QuestionAnswered = async ({
 }) => {
   try {
     const result = await generateObject({
-      model: DEFAULT_MODEL,
+      model: createModel(DEFAULT_MODEL),
       system: `You are an AI evaluation expert. Your job is to determine if an AI assistant actually answered the customer's question with a specific answer, or if it deflected, explained why it couldn't answer, or gave generic suggestions without providing the requested information.
 
 Instructions:

package.json

@@ -31,6 +31,7 @@
     "@composio/vercel": "^0.3.4",
     "@fal-ai/client": "^1.9.5",
     "@modelcontextprotocol/sdk": "^1.24.3",
+    "@openrouter/ai-sdk-provider": "^2.6.0",
     "@privy-io/node": "^0.6.2",
     "@supabase/supabase-js": "^2.86.0",
     "@trigger.dev/sdk": "^4.4.3",