Seph Gentle delivered a closing talk at a local-first software conference, arguing that the future of interoperable applications depends on a more robust and sustainable data format. He framed the current landscape as a false choice between traditional local software, which remains confined to a single device, and cloud-based software, which often holds user data hostage within centralised services. To overcome this, Gentle proposed a conceptual local-first operating system where programs share a common data layer, allowing them to evolve independently while remaining fully compatible.
A central theme of the talk was the necessity of a durable schema format. Gentle asserted that a high-quality format must be self-describing, statically typed, and capable of maintaining compatibility over long time horizons. Furthermore, it should preserve foreign data that it does not yet understand and be optimised for efficient network synchronisation and data streaming.
To achieve these goals, Gentle introduced the concept of embedding the schema directly into the file header. This approach allows any tool to inspect and interpret data without requiring a separately distributed schema file. He contrasted this with Protocol Buffers or Cap’n Proto, which rely on pre-shared schemas, and JSON, which he characterised as being too verbose and weakly typed for long-term data durability.
He illustrated these design principles using an experimental Rust implementation called schemaboi. In this system, the schema is stored in the initial bytes of every file and is merged with the application's internal schema upon opening. The type system supports primitives such as booleans, integers, and floats, as well as complex structures such as lists, maps, and a unified struct-and-enum model. Critically, missing fields are filled with defaults, while unknown fields are preserved rather than discarded, ensuring that data integrity is maintained even when software versions differ.
Gentle also established clear definitions for different types of compatibility. Beyond standard backward compatibility (opening old files in new apps) and forward compatibility (opening new files in old apps), he introduced "sideways compatibility." This allows disparate developers to add fields independently without central coordination. To facilitate this, he advocated for globally unique field identifiers instead of simple incrementing tags, thereby preventing collisions during parallel development.
The talk included specific safety guidelines for schema evolution. Gentle noted that adding optional fields or new enum variants is generally safe. Conversely, making a field mandatory, renaming fields, or altering a field's type were flagged as high-risk actions. He emphasised a preference for simple, incremental enhancements, citing the enduring success of systems like email and HTTP as evidence that evolution is often preferable to forced migrations.
In his conclusion, Gentle addressed practical considerations such as canonical encoding for content-addressable systems and the use of ATProto-style `must_use` fields. While benchmarks showed that his format is significantly smaller and faster than JSON, he cautioned that the project remains experimental. Ultimately, his vision points toward a future where data is truly owned by the user, independent of the specific applications that created it.