How it works

microphone + call  →  transcript (on your Mac)  →  question?  →  search  →  answer  →  above the pill

Smalt thinking

1. Each new line goes to the question spotter: TypeSafe Jev if you added its key, otherwise the cloud model.
2. For a real question, Smalt searches your folders and keeps the best few passages.
3. The cloud model writes a short answer from them, or from general knowledge when your folders don't cover it.

Your folders are indexed once into a small file on your Mac (~/Library/Application Support/Smalt/knowledge.sqlite). It records which words are on which lines, not the text itself, and only new or changed files are read again, a couple of seconds after you save them. A search takes about a tenth of a second, and nothing is held in memory between questions.

Claude Code thinking

Smalt runs a small local server (MCP) that only apps on your Mac can reach. In a meeting, Claude waits for new lines, decides what deserves an answer, searches with its own tools and sends the answer back. Smalt keeps no keys and calls no model in this mode.

Talking

With /smalt, Smalt hears you with the speech model on your Mac and hands each turn to Claude Code. Claude's reply is said by a voice model on your Mac. A hook in Claude Code tells Smalt which step runs, so the pill shows it. No audio leaves your Mac. More in Talk with Claude Code.

Where things live