Mozilla.ai is building Otari as a gateway that lets applications use open models without giving up the conveniences normally bundled into proprietary AI APIs. It targets the less visible services around a model, including web search, code execution, context handling and persistent memory.
With a hosted assistant, a search query or uploaded file may be processed by infrastructure chosen by the provider. Otari instead routes search through services the developer configures with their own credentials and runs code in specified sandbox environments. The application sees a consistent tool interface even when the underlying model host or service changes.
Mozilla.ai says it has already used the approach in an agent application, but acknowledges that work remains. The project still needs further development around context compaction, prompt caching and memory that can be inspected and moved between models.
The benchmark Mozilla.ai has set is demanding: an agent designed for a commercial frontier API should be able to point at open models through Otari and keep working. Otari has not yet met that goal in every case, but it provides a concrete architecture for teams that value provider choice and control over where agent tools execute.