SHEV is a minimal decision-maker that consumes one whole message (an email, ticket, review, etc.) and returns a single, structured JSON decision containing an actionable string, a boolean take_action, a numeric confidence, and a generated_response. It deliberately avoids agents, orchestration, long prompt engineering, or line-by-line parsing: one message in, one judgment out. The interface supports either a small local ternary model (Bonsai) running on Apple Silicon via MLX or a remote DeepSeek backend via API key; the local default is Bonsai, while DeepSeek runs over the network and stores its key in ~/.config/shev/config.json. Decisions are appended to responses.json so other tools can read the records.
Installation offers a one-line curl installer, an npm package (@noorslvt/shev) with caveats about registry names and install scripts, or cargo install from the repo; uninstall cleans binaries, keys, local models, and tracked chat records. The terminal UI shows model output and the raw JSON for the current decision side-by-side, with commands for loading files, managing models/keys, copying JSON, history, MCP bridge/tools, and quick keybindings (enter to send, ctrl+t to switch model). Recommended hardware is Apple Silicon (M1 or newer, 16 GB ideal), though DeepSeek lets it run on any Rust-supported platform.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.