Capabilities
Performance
Every metric is sharpened so multi-session workflows stay smooth even on a laptop. Below are PSS measurements from the README.
RAM (1 active session)
| Tool | Resident set |
|---|---|
minnal (embeddings off) | 27.8 MB — baseline |
minnal (embeddings on) | 167.1 MB |
pi | 144.4 MB |
Codex CLI | 140.0 MB |
GitHub Copilot CLI | 333.3 MB |
OpenCode | 371.5 MB |
Startup
- Cold start:
< 100 msto first frame. - Time to first input: under one frame on most terminals.
- Render rate:
> 1000 fps; capped by your monitor.
Scaling additional clients
Memory growth per attached client is sub-linear because the server holds the heavy state (memory graph, session history, embeddings). Adding a second TUI client costs a few MB, not another full process.
Design choices that buy this
- Single static Rust binary — no Node, no Python, no Electron.
- Lazy loading of providers, MCP servers, skills.
- Custom mermaid renderer in Rust (~1800× faster than the JS version).
- Custom terminal (
handterm) for partial-line scrolling without flicker. - Streaming-first I/O for sub-100 ms time-to-first-frame.