We just open-sourced a small customization package that adds OpenRouter (Direct) as a provider on Agent LLM Connections (ML201000) in 2026 R2 (built against 26.200.0334). If you haven’t used OpenRouter, it’s a single API and a single bill in front of hundreds of models from OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek and others. That makes it easy to put a cheap, fast model on classification agents and a stronger one on summaries, compare them on real records, and switch when something better or cheaper ships, all without separate vendor accounts. It can also route around a provider that’s down, and your account settings control which upstream providers see your data. Out of the box, AI Automation agents reach models through Acumatica’s cloud proxy and its fixed provider list. The platform also has a hidden “direct provider” path that builds the chat client inside the site; the only one shipped is Azure (Direct). This package registers another direct provider that points the site’s own OpenAI client at OpenRouter’s OpenAI-compatible endpoint. Models are set per connection with no code change, and it works on sites without an AI Studio proxy subscription. Everything above the transport stays stock: instructions, tools, field masking, token counts in AI Automation History, and the Test Agent diff and processing log.
Setup takes three steps. First, import and publish the zip on SM204505; it’s a single DLL with no screens or DAC changes. Second, add <ai-studio EnableDirectProviders="true" /> inside <px.core> in web.config, which a package can’t do for you. Third, create a connection with your OpenRouter API key and a model id such as openai/gpt-4o-mini, then run Test Connection. Pick models that support tool calling and structured outputs, because agents rely on both. A few caveats apply. The interfaces it hooks are marked [PXInternalUseOnly], so re-test after every upgrade and don’t expect ISV certification. The proxy’s 800-token default output cap doesn’t apply. Your data also goes to OpenRouter and the model’s upstream provider, so check that against your data-handling requirements. The same two-class pattern works for any OpenAI-compatible endpoint, such as vLLM, LM Studio, Groq or Together.
The package is MIT licensed: https://github.com/contou-consulting/acumatica-aistudio-open-router-provider
