Gorkie uses multiple AI model providers with automatic fallback to ensure reliability. The provider system is built on Vercel AI SDK with custom retry logic.
Provider Configuration
Providers are configured in server/lib/ai/providers.ts.
Available Providers
Hack Club AI Proxy
Primary provider using Hack Club’s AI proxy service.
const hackclubBase = createOpenRouter({
apiKey: env.HACKCLUB_API_KEY,
baseURL: 'https://ai.hackclub.com/proxy/v1',
});
const hackclub = wrapProvider({
provider: hackclubBase,
languageModelMiddleware: {
specificationVersion: 'v3',
overrideProvider: () => 'hackclub',
},
imageModelMiddleware: {
specificationVersion: 'v3',
overrideProvider: () => 'hackclub',
},
});
OpenRouter
Fallback provider for reliability.
const openrouter = createOpenRouter({
apiKey: env.OPENROUTER_API_KEY,
});
Model Selection
Chat Model
Main conversational model with multiple fallbacks.
const chatModel = createRetryable({
model: hackclub.languageModel('google/gemini-3-flash-preview'),
retries: [
hackclub.languageModel('google/gemini-2.5-flash'),
hackclub.languageModel('openai/gpt-5-mini'),
openrouter('google/gemini-3-flash-preview'),
openrouter('google/gemini-2.5-flash'),
openrouter('openai/gpt-5-mini'),
],
onError: onModelError,
});
- Hack Club →
google/gemini-3-flash-preview
- Hack Club →
google/gemini-2.5-flash
- Hack Club →
openai/gpt-5-mini
- OpenRouter →
google/gemini-3-flash-preview
- OpenRouter →
google/gemini-2.5-flash
- OpenRouter →
openai/gpt-5-mini
Summarizer Model
Lightweight model for thread summarization.
const summariserModel = createRetryable({
model: hackclub.languageModel('google/gemini-3-flash-preview'),
retries: [
hackclub.languageModel('google/gemini-2.5-flash'),
hackclub.languageModel('openai/gpt-5-mini'),
openrouter('google/gemini-2.5-flash-lite-preview-09-2025'),
openrouter('openai/gpt-5-nano'),
],
onError: onModelError,
});
- Hack Club →
google/gemini-3-flash-preview
- Hack Club →
google/gemini-2.5-flash
- Hack Club →
openai/gpt-5-mini
- OpenRouter →
google/gemini-2.5-flash-lite-preview-09-2025
- OpenRouter →
openai/gpt-5-nano
Image Model
For image generation tasks.
export const provider = customProvider({
languageModels: {
'chat-model': chatModel,
'summariser-model': summariserModel,
},
imageModels: {
'image-model': hackclub.imageModel('google/gemini-3.1-flash-image-preview'),
},
});
Error Handling
The onModelError callback logs failures and automatically switches to the next model.
const onModelError = (context: {
current: { model: { provider: string; modelId: string } };
}) => {
const { model } = context.current;
logger.error(
`error with model ${model.provider}/${model.modelId}, switching to next model`
);
};
Provider Setup
Environment Variables
Required environment variables (defined in server/env.ts):
HACKCLUB_API_KEY=your_hackclub_api_key
OPENROUTER_API_KEY=your_openrouter_api_key
Usage in Code
Import the provider and reference models by name:
import { provider } from '~/lib/ai/providers';
import { streamText } from 'ai';
const result = await streamText({
model: provider.languageModel('chat-model'),
system: systemPrompt,
messages,
tools,
});
Custom Provider
The final exported provider is a custom provider that wraps the chat, summarizer, and image models:
export const provider = customProvider({
languageModels: {
'chat-model': chatModel,
'summariser-model': summariserModel,
},
imageModels: {
'image-model': hackclub.imageModel('google/gemini-3.1-flash-image-preview'),
},
});
Retry Logic
The createRetryable function from ai-retry automatically:
- Tries the primary model first
- Falls back to each retry model in order on error
- Logs each failure via
onModelError
- Throws if all models fail
This ensures maximum uptime even if individual providers or models experience issues.
Model Selection Strategy
Performance vs. Cost: The primary models (Gemini 3 Flash Preview) offer the best balance of speed and quality. Fallbacks include both newer (Gemini 2.5) and lighter (GPT-5 Mini/Nano) models for redundancy.
API Key Security: Never commit API keys to version control. Use environment variables and ensure they’re listed in .gitignore.