BoxTalk

Local text-to-speech on your desktop, paired with a Chrome extension that pipes any web page (or selected text) to it. Powered by Kokoro — an 82M-parameter open-weights TTS model — running entirely on your machine via onnxruntime-node. No cloud round-trip, no API key, no telemetry. Ships with an optional second engine (VibeVoice) for higher-fidelity playback and voice cloning; see Engines below. English only for now; the models support other languages and they're on the roadmap.

Why

Most "read this article aloud" tools either ship audio to a cloud API or use the OS's built-in voices. BoxTalk does neither: it runs Kokoro locally so your reading list never leaves your laptop, and the voices are dramatically better than say / SAPI / espeak. The Chrome extension is the glue — you keep browsing, you queue or speak from the page you're on, and BoxTalk handles synthesis, playback, and history.

Two pieces

┌──────────────────────────┐                    ┌───────────────────────────┐
│  Chrome extension (MV3)  │  HTTP + token      │   BoxTalk desktop app     │
│  • right-click selection │ ─────────────────▶ │   • Kokoro TTS engine     │
│  • whole-page popup      │  127.0.0.1:38219   │   • Queue + History       │
│  • Cmd/Ctrl+Shift+S      │                    │   • Local SQLite store    │
└──────────────────────────┘                    └───────────────────────────┘

Both halves, plus the marketing site, live in this monorepo:

apps/
  desktop/     Electron + React desktop app (TTS engine + UI)
  extension/   MV3 Chrome extension (selection / page → desktop bridge)
  sites/
    landing/   Astro marketing site (boxtalk.dev)

Install

1. Get the desktop app

Pre-built release (recommended) — grab the latest DMG from the Releases page, drag BoxTalk.app into Applications, and launch it. First launch downloads the quantized Kokoro model (~80 MB) from Hugging Face; every launch after that runs offline.

From source — if you want to hack on the app or there's no release for your platform yet:

pnpm install
pnpm start        # builds the renderer and launches Electron

You'll need Node 20+ and pnpm. To produce a distributable DMG:

pnpm dist:mac     # writes apps/desktop/release/*.dmg

2. Load the Chrome extension

The extension isn't on the Chrome Web Store yet — load it as an unpacked extension:

1. Open chrome://extensions and toggle Developer mode on.
2. Click Load unpacked and pick apps/extension/ from this repo.
3. In the BoxTalk app, open General → Browser extension and copy the pairing token.
4. Click the extension's toolbar icon → Settings → paste the token → Save.

The popup's status dot should turn green when it can reach the desktop app.

How the two halves talk

The desktop app runs a minimal HTTP server on http://127.0.0.1:38219 — loopback only, never bound to a public interface. Every mutating request must carry an X-BoxTalk-Token: <pairing-token> header. The token is a 24-byte random string generated on first launch and stored in the app's local SQLite database. It surfaces in General → Browser extension so you can paste it into the extension's settings page.

The bridge exposes three endpoints:

Endpoint	Auth	What it does
GET /status	none	Health check. Returns { ok, kokoro } so the extension can render its connection dot.
POST /speak	token	"Read this now." The app focuses, chunks the text sentence-by-sentence, and plays it.
POST /candidates	token	"Save this for later." Inserts a row into the Queue (Waiting column). The app stays quiet.

The extension never opens an arbitrary host. Its only host permission is http://127.0.0.1:38219/*. Without the desktop app running on loopback nothing happens — there's no fallback path.

Using the extension

There are two modes and two actions, which cross-multiply to four flows. All of them post text from the active tab to the desktop app.

• Selection — highlight text on any page, then either right-click → "Speak with BoxTalk", press Cmd/Ctrl+Shift+S, or open the popup's Selection tab.
• Whole page — open the popup's Whole page tab. The extension runs Mozilla Readability on a clone of the live DOM (same algorithm as Firefox Reader View) and shows the article body so you can edit or trim before sending.

For each mode you can choose:

• Speak now → /speak. The desktop app jumps to the front and starts reading.
• Save for later → /candidates. Lands silently in the Queue with the source URL and page title.

Long pages are chunked by the desktop app and played back-to-back — each chunk is also written to the History tab tagged with the source URL.

The Queue

The Queue is the first screen you land on, because it's where saved-for-later pages live. Pick the voice for new digests at the top of the page — the choice is persisted, and "Heart" is the default. Hit Digest on a waiting card and the desktop app chunks the text, synthesizes each chunk through Kokoro, merges the Float32 samples into one WAV, and writes the file to the app's user-data dir (userData/digests/<id>.wav). Playback after that is instant and works without the model loaded.

Once digested, a card moves to the right-hand Digested column and reveals a self-contained transport panel:

• Skip back 30s · Play / Pause · Skip forward 30s — circular icon buttons.
• A draggable seek bar with a hover tooltip that surfaces the m:ss you'd jump to. The played portion fills in pink as audio progresses.
• A playback-speed segmented control (0.75× · 1× · 1.25× · 1.5× · 2×). The chosen speed drives audio.playbackRate live — no re-synthesis — and the setting persists across plays. Digest itself always renders at 1.0×; speed is a listening preference, not a synthesis parameter.

The other views

View	What it does
Queue	Pages + selections from the extension. Pick the voice at the top, Digest to render, then a full transport per card.
General	App-level settings: selected model, history retention, and the pairing token.
Model	Live model status. Kokoro 82M loads automatically on first launch (~80 MB download, then cached).
History	Every synthesis recorded, with text, voice, source, timestamp, duration. Live search + clear.
Speak	Free-form text input. Type, pick speed, hit Speak.

Engines

BoxTalk ships with two TTS engines. Kokoro is on by default; VibeVoice is opt-in.

Kokoro 82M (default)

82M-parameter open-weights model from hexgrad/Kokoro, running through kokoro-js + onnxruntime-node. Quantized to q8 ONNX, ~80 MB download, comfortable on a laptop CPU. Ships with 13 English voices — American + British, female + male. Listed in the Model card; click the play button on any row to preview the voice, the radio dot to make it the default.

VibeVoice (optional, higher fidelity)

vibevoice.cpp — a C++ port of Microsoft VibeVoice that runs on CPU via GGUF weights. Three model sizes are surfaced in the UI:

Model	Size	What you get	Status
VibeVoice Realtime 0.5B	~2 GB	Pre-baked Carter + Emma voices, low-latency streaming	Supported — GGUFs published on Hugging Face, downloads on demand
VibeVoice 1.5B	~6 GB	Up to 4 speakers, voice cloning from reference audio	Requires manual conversion of upstream weights
VibeVoice Large 7B	~14 GB	Highest fidelity, ModelScope mirror	Requires manual conversion of upstream weights

VibeVoice is entirely on-demand. Nothing about it ships in the DMG. The first time you click Load on VibeVoice Realtime 0.5B in the Model tab, BoxTalk pulls two things into the app's user-data dir:

1. vibevoice-cli — a prebuilt binary for your platform+arch, downloaded from the matching BoxTalk release asset (e.g. vibevoice-cli-darwin-arm64). Cached at <userData>/vibevoice/bin/vibevoice-cli.
2. The GGUF weights (~2 GB) — pulled from mudler/vibevoice.cpp-models on Hugging Face. Cached at <userData>/vibevoice/models/.

If no prebuilt is published for your platform yet (or you're offline), BoxTalk falls back to cloning + compiling localai-org/vibevoice.cpp locally — requires Xcode Command Line Tools or your platform's equivalent (git, cmake, make).

To build the CLI manually for development:

pnpm build:vibevoice   # clones localai-org/vibevoice.cpp into vendor/ and builds

The 1.5B and 7B cards show a friendly error until you point them at converted weights (see apps/desktop/vibevoice.js).

Tests

pnpm smoke    # all 13 voices generate a WAV
pnpm e2e      # full UI flow via Playwright + Electron
pnpm test     # smoke + e2e

See apps/desktop/README.md for what each test covers, and apps/extension/README.md for loading + debugging the extension.

Roadmap

• Chrome extension that captures selected text / page content and pipes it to this app.
• Queue view with a per-card digest-then-listen workflow.
• Always-visible playback transport on digested cards: pause/resume, ±30s skip, draggable seek with hover tooltip, live playback-speed control.
• Optional second engine: VibeVoice (vibevoice.cpp) — Realtime 0.5B working today, 1.5B + 7B pending GGUF availability.
• More languages — Kokoro supports Japanese, Mandarin, Spanish, French, Hindi, Italian, Portuguese.
• "Summarize then read" mode driven by a local LLM in the extension.

Landing site

The marketing site lives at apps/sites/landing/ (Astro + Tailwind). It deploys to Cloudflare Pages as project boxtalk — currently reachable at https://boxtalk.pages.dev/.

.github/workflows/deploy-landing.yml redeploys whenever:

• apps/sites/landing/** (or the lockfile / workflow file) changes on main
• A GitHub release is published — so the landing's build-time fetch of releases/latest picks up the new version + DMG link

For CI deploys you need two repo secrets (Settings → Secrets and variables → Actions):

Secret	Value
CLOUDFLARE_API_TOKEN	Create at https://dash.cloudflare.com/profile/api-tokens with the Cloudflare Pages — Edit template (or a custom token with Account.Cloudflare Pages: Edit), scoped to the Accomplish account.
CLOUDFLARE_ACCOUNT_ID	ab43a89c284963fce47460305a945611 (run wrangler whoami to reconfirm)

Deploy manually from a local checkout with:

pnpm -F @boxtalk/site build
npx wrangler pages deploy apps/sites/landing/dist --project-name=boxtalk --branch=main

To attach the boxtalk.dev custom domain, first add the zone to Cloudflare (Add Site flow), then in the Pages dashboard go to boxtalk → Custom domains → Set up a custom domain → boxtalk.dev.

Releasing

Tag pushes trigger the macOS release build via .github/workflows/release.yml:

# Bump version in apps/desktop/package.json first, then:
git tag v0.1.0
git push origin v0.1.0

CI runs on macos-latest, builds the unsigned DMG via pnpm -F @boxtalk/app dist:mac, and uploads it to a GitHub release named after the tag. The landing site (apps/sites/landing) reads releases/latest at build time and links the DMG directly. To dry-run the build without publishing, use the workflow's workflow_dispatch trigger.

The build is unsigned (identity: null in electron-builder config), so first-launch users need to right-click the app → Open, or run xattr -dr com.apple.quarantine /Applications/BoxTalk.app.

License

MIT.