MacPaw and Liquid AI are building an on-device AI stack for Mac apps

MacPaw and Liquid AI are co-developing an on-device AI stack for macOS, starting with MacPaw's Eney assistant. For AI builders, the important shift is architectural: this is intended to combine a local model, memory, inference, and app context into a reusable Mac platform rather than simply adding a cloud API to an existing product.

Eney is the first test of a local Mac AI layer

The partnership brings Liquid AI LFMs together with MacPaw's Mnemos memory system and Elix inference technology. MacPaw says Eney already uses local intelligence for reasoning, contextual search, skill execution, and conversation history where possible. The new work is meant to extend that approach through joint stack development.

That distinction matters. MacPaw is not describing a plug-in model swap. It and Liquid AI are adapting the underlying components for Mac-based assistance, then testing the result in a product with an existing assistant workflow.

The practical opportunity is app-aware assistance

The proposed value is not merely that inference happens locally. After Eney is tested, MacPaw plans to extend the technology to more products and make the shared stack available to Mac developers through Setapp. Participating apps could expose more of their capabilities and context to Eney, allowing the assistant to answer questions about app data or take relevant actions inside an app.

For builders, that points to a local AI layer connecting models with application permissions, memory, and action APIs. A task such as finding information, summarizing local work, or triggering an app function could require less data movement and fewer round trips to a remote service. The quality of that experience will depend on the integration contract and permission model, neither of which has been fully detailed.

Local inference helps, but it does not solve every product problem

On-device AI for Mac can reduce dependence on cloud inference and may keep more reasoning and conversation history on the computer. That can be useful for privacy-sensitive workflows, offline availability, and products where recurring cloud inference costs are difficult to control.

However, local inference also puts constraints on model size, memory use, thermal load, and performance across different Macs. The announcement does not provide independent benchmarks, hardware support details, latency figures, or a complete data handling policy. Builders should therefore treat the privacy and performance benefits as design goals, not verified product outcomes.

What Mac developers should watch next

Setapp access is described as a future step after Eney testing and broader evaluation, not an immediately available SDK. Developers should watch for documentation covering supported macOS versions, Apple silicon requirements, app data permissions, model updates, memory persistence, fallback behavior, and whether local and cloud execution can be mixed.

The strongest near-term conclusion is limited but useful: MacPaw and Liquid AI are building toward a shared local AI runtime for Mac apps. It becomes strategically important for developers only if Eney demonstrates reliable app actions and Setapp provides a clear way to connect those actions to third-party software.