NeuroLink gives an application one API for every kind of AI provider: OpenAI, Anthropic, Google, AWS, Azure, Mistral, local runtimes and dozens more, switchable with a single parameter change, alongside a model registry, MCP support and multi-provider failover.
Alongside its generate and stream calls for text, a decide call asks Jev for a typed, calibrated judgment instead: a boolean, a choice or a score. Five call sites inside NeuroLink itself use it this way, for model routing, deciding which earlier messages still matter, checking a summary kept every open question, and picking which tools a request actually needs.
Extracted from production systems at Juspay, where it also powers three of the company's own products; every Jev-backed decision fails open, so nothing changes in its absence if no key is configured.
What you get
- Full TypeScript source code under the MIT license
- The provider abstraction layer, the decide inference type, and its five built-in call sites






