Tiles: Own your AI with local models and open protocols

As AI becomes central to how we work and communicate, I believe we’re also seeing a new form of technological feudalism, where a handful of platforms control the models, identities, data, and distribution channels people depend on. Local-first AI offers an alternative, but many tools still lack the identity, sync, and collaboration features people expect.

In this talk, I’ll share what I’ve learned building Tiles, a local-first private AI assistant. I’ll discuss the tradeoffs I’ve encountered while combining local models with decentralized identity, peer-to-peer encrypted sync, and AT Protocol–based collaboration.

I’ll also cover Tiles’ deliberately opinionated approach of deeply integrating a single model and agent harness at a time. Today, that means pairing gpt-oss-20b with the Pi harness. I’ll share the reasoning behind this design choice and how it shapes the product’s capabilities and constraints.

I’ll also present some of the exploratory research we’re conducting to bring the convenience of modern AI services to a user-owned stack. This includes ongoing work around remote agents, sandboxed extensions, and lexicon-based automation. I’ll discuss how these ideas could enable more interoperable, collaborative, and decentralized AI systems.