Documentation¶

Guides for using localvoxtral, its coding-agent integrations, and working on its code.

Using localvoxtral¶

Install: install, update, nightly builds
Dictating: shortcuts, output modes, settings reference, screenshots
Terminals & coding agents: dictating into Claude Code and other CLI agents, session joins, the SSH remote plugin
Work with several agents: being told an agent needs you, jumping to it, and naming, reaching and sending to sessions by voice
Integration matrix: for each coding agent and terminal, what joins, what context is attached, and why the gaps exist
Under the hood: privacy, the managed local engines and their pinned models, bring-your-own-server
Roadmap

Claude Code
opencode
Mistral Vibe
Codex
herdr: dictating into a herdr pane, on this Mac, on an ssh host and on a federated machine
Remote Claude Code over SSH: enrolling a host, what the token does and does not authorize, the per-Mac forward port, manual checks, and uninstalling

Building from source, plus CONTRIBUTING.md for contribution expectations
Architecture: the subsystem map
Test harness: the control socket and the WAV microphone that debug and UI smoke builds carry, and release builds never

These live in the agent folder of the docs. Agents load them on demand from AGENTS.md, and people can read them too.

Invariants & deliberate tradeoffs: trust boundaries, session-join arms, fail-closed rules. Read before touching the Claude Code context path.
Test tiers & eval lanes: the full tier matrix, when the LLM lanes must run, eval recordings and ablations
Field debugging: try-pr, crashlog dispatch, signing/TCC, diagnostic records

Machine-local scratch (setup runbooks, handoff notes, drafts) goes in the gitignored local-notes/ directory, never in the docs folder.