I run a 'radar' that finds free LLM endpoints and auto-adopts the good ones — behind a five-part gate so it can't adopt junk

Free LLM endpoints appear and disappear frequently due to various reasons, such as providers launching generous tiers, new open models surfacing on hosts' gateways, or promotional campaigns running for limited periods. Manually tracking them is a futile endeavor. This prompted the creation of a small radar (model-radar) designed to monitor for free models and automatically adopt those that prove to be genuinely good, provided they pass a rigorous gate screening process.

This gate consists of five sequential checks that a model must pass before it can be adopted into the agent stack. The radar is designed to prevent the adoption of poor-quality models that could lead to silent failures and degrade the performance of agent tasks. The gate verifies, in order, whether the model responds correctly within a reasonable timeout, can perform tool calls accurately and incorporate the results into subsequent turns, respects prompt context, maintains stability across repeated calls, and outperforms the existing models on specific agent tasks.

Only a model that successfully passes all five checks is allowed to be adopted. The radar's design ensures that the automated adoption process remains transparent and reproducible, minimizing the risk of introducing silent failures. By gating the adoption of free models based on their ability to handle real tasks, model-radar aims to prevent the adoption of suboptimal models that may appear functional on surface-level metrics but fail to deliver desired outcomes in actual agent deployments.