The voice of Claude Code — a native macOS app that reads replies aloud and takes dictation, fully on-device.
flowcode is a tiny native macOS menu-bar app that gives Claude Code a voice. The moment a new reply lands, it reads the prose aloud with a local text-to-speech engine — so you hear when Claude finishes, asks a question, or needs your call, without babysitting the terminal.
When it's your turn, you hold Right Option (⌥) and talk; your words are transcribed locally and pasted into the focused app — and you press Enter. It also reads the Claude desktop app (Chat, Cowork, Code). Everything runs on-device: no account, no cloud, no telemetry.
Living in Claude Code for long sessions means two kinds of waiting: watching the terminal for a reply to finish, and typing the next prompt when your hands or eyes are busy. Voice for coding agents has been built many times — but it tends to feel turn-based and clunky, a loop bolted onto a terminal rather than a finished product.
I wanted something that felt ambient: Claude simply speaks when it has something to say, and I can answer out loud without breaking flow — while staying completely private and never letting a voice command do anything irreversible on its own.
flowcode never touches Claude Code's internals. Claude Code already writes a live JSONL transcript of every session to disk — flowcode just tails that file, extracts each new assistant message, strips tool-call narration and code blocks so only the prose is spoken, and sends it to a local Kokoro TTS service. For the Claude desktop app, which has no transcript file, it reads the on-screen text through the macOS Accessibility tree instead.
Dictation is the mirror image: holding ⌥ captures the mic, a local Whisper service transcribes it, and the text is pasted into whatever app is focused. The mic is released the instant you let go. Both engines are localhost services, so nothing ever leaves the Mac.
Claude Code JSONL ─tail─┐
├─▶ flowcode ─HTTP─▶ Kokoro :8880 ─▶ 🔊 + orb
Claude Desktop AX ─────┘
⌥ hold ──▶ mic ──HTTP──▶ Whisper :2022 ──paste──▶ focused appThe product is split into two tiers. Model B — what ships today — is read-aloud plus dictation, written entirely in Swift with no Python core and no socket: fast, small, dependency-light. Model A is an experimental real-time layer (barge-in, streaming TTS, semantic endpointing) that runs a forked voicemode Python core; it is wired in but default-off.
Two principles are non-negotiable. First, everything is local — TTS and STT are localhost services, there is no account and no network round-trip. Second, voice can only propose, never commit: flowcode never presses Enter for you. Higher-risk actions go through a confirmation gate that requires a real, non-voice gesture (a click, a hotkey, or Touch ID), so ambient sound, a colleague's voice, or the app's own TTS can never trigger anything — and a timeout always defaults to deny.
flowcode is open source under MIT and in public beta (v0.3.0). It installs in an unusual way — agent-driven: you point Claude Code at the repo and it runs the setup end-to-end (voice services, build, signing, permissions). The next big lever is a notarized one-click download, which is what would take it from a few hundred build-from-source users to a broad audience. Built on top of voicemode by Mike Bailey.