128mb local-first desktop ai agent

The ERTH stack ships as a single 128MB binary that combines an ElectroBun UI shell, a Robyn-powered Python sidecar, and a Turso libSQL database, all orchestrated by HTMX [DevTo]. ElectroBun replaces Electron’s Chromium engine with the native WebKit implementation and runs on Bun, achieving launch times under 200ms [GitHub]. The Python sidecar runs Robyn, a Rust-backed async framework, and communicates with the UI via a local HTTP channel that uses UUIDv7 opaque tokens for authentication. Turso stores data in a file-based libSQL database; benchmarked reads hit 0.1ms latency on a typical SSD. The entire stack is open-source on GitHub and can be built for macOS, Windows, and Linux.

The ERTH binary stays under 128MB and idles below 50MB, making it viable for low-spec machines and corporate BYOD policies. By bundling a Python sidecar that can load local LLM weights, developers can ship AI assistants that run entirely offline, sidestepping data-privacy concerns and recurring cloud costs [DevTo]. The architecture lets teams use Bun for rapid UI iteration and Python for AI logic without committing to Rust’s steep learning curve, preserving developer velocity while still delivering native performance [GitHub].