Add agent-voice: local push-to-talk dictation for coding agents

.github/workflows/agent-voice.yml

@@ -0,0 +1,103 @@
+name: agent-voice
+
+# agent-voice is a standalone package (own bun.lock, not a workspace member), so
+# it gets its own workflow instead of riding along with the monorepo CI.
+#
+# The Rust leg runs `cargo check`/`clippy`/`test` rather than a full Tauri
+# bundle: building the app needs WebView2, CMake and LLVM, which are slower and
+# unavailable on hosted runners. `cargo check --tests` still compiles the
+# `tauri::generate_context!` macro, so configuration errors (capabilities,
+# window labels, updater key) fail here.
+
+on:
+  push:
+    branches: [main]
+    paths:
+      - 'Apps/agent-voice/**'
+      - '.github/workflows/agent-voice.yml'
+  pull_request:
+    paths:
+      - 'Apps/agent-voice/**'
+      - '.github/workflows/agent-voice.yml'
+
+concurrency:
+  group: agent-voice-${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+defaults:
+  run:
+    working-directory: Apps/agent-voice
+
+jobs:
+  typescript:
+    name: TypeScript (app + worker)
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version-file: .bun-version
+
+      - name: Install dependencies
+        run: bun install --frozen-lockfile
+
+      - name: Typecheck
+        run: bun run typecheck
+
+      - name: Test (frontend helpers + license worker)
+        run: bun test
+
+      - name: Build frontend
+        run: bun run build
+
+  rust:
+    name: Rust core
+    runs-on: ${{ matrix.os }}
+    timeout-minutes: 45
+    strategy:
+      fail-fast: false
+      matrix:
+        os: [ubuntu-latest, windows-latest]
+    steps:
+      - uses: actions/checkout@v6
+
+      - uses: oven-sh/setup-bun@v2
+        with:
+          bun-version-file: .bun-version
+
+      - uses: dtolnay/rust-toolchain@stable
+        with:
+          components: rustfmt, clippy
+
+      - name: Install Linux build dependencies
+        if: runner.os == 'Linux'
+        run: |
+          sudo apt-get update
+          sudo apt-get install -y libwebkit2gtk-4.1-dev libappindicator3-dev \
+            librsvg2-dev patchelf libxdo-dev cmake clang libssl-dev
+
+      - name: Cache cargo
+        uses: Swatinem/rust-cache@v2
+        with:
+          workspaces: 'Apps/agent-voice/src-tauri -> target'
+
+      - name: Install frontend dependencies
+        run: bun install --frozen-lockfile
+
+      - name: Frontend build (tauri.conf.json points at ../dist)
+        working-directory: Apps/agent-voice
+        run: bun run build
+
+      - name: Formatting
+        working-directory: Apps/agent-voice/src-tauri
+        run: cargo fmt --check
+
+      - name: Clippy
+        working-directory: Apps/agent-voice/src-tauri
+        run: cargo clippy --all-targets -- -D warnings
+
+      - name: Tests
+        working-directory: Apps/agent-voice/src-tauri
+        run: cargo test

Apps/agent-voice/.env.example

@@ -0,0 +1,19 @@
+# Paddle checkout overlay. Copy to `.env` and fill in — `.env` is gitignored.
+#
+# Client-side token from Paddle → Developer tools → Authentication. It is safe
+# to ship in the bundle; it only authorises opening checkout.
+VITE_PADDLE_TOKEN=
+
+# Price id of the one-time Pro price (Paddle → Catalog → product → Prices).
+VITE_PADDLE_PRICE_ID=
+
+# `sandbox` (default) or `production`.
+VITE_PADDLE_ENV=sandbox
+
+# Build-time Rust knobs. These are *not* read from .env — pass them to cargo:
+#
+#   AGENT_VOICE_WORKER_URL=https://licenses.example.com \
+#   AGENT_VOICE_ENTITLEMENT_PUBKEY="$(base64 -w0 entitlement.pub)" \
+#     cargo tauri build
+#
+# See docs/entitlement.md before rotating either one.

Apps/agent-voice/.gitignore

@@ -0,0 +1,31 @@
+# Logs
+logs
+*.log
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+pnpm-debug.log*
+lerna-debug.log*
+
+node_modules
+dist
+dist-ssr
+*.local
+
+# Editor directories and files
+.vscode/*
+!.vscode/extensions.json
+.idea
+.DS_Store
+*.suo
+*.ntvs*
+*.njsproj
+*.sln
+*.sw?
+
+# Rust build output and generated Tauri schemas
+src-tauri/target
+src-tauri/gen/schemas
+
+# Wrangler state
+worker/.wrangler

Apps/agent-voice/.vscode/extensions.json

@@ -0,0 +1,3 @@
+{
+  "recommendations": ["tauri-apps.tauri-vscode", "rust-lang.rust-analyzer"]
+}

Apps/agent-voice/README.md

@@ -0,0 +1,122 @@
+# agent-voice
+
+Hold a hotkey. Speak. The text lands in whatever you were typing — locally,
+with no audio ever leaving the machine.
+
+Built for people who dictate to coding agents: hold `Ctrl/Cmd+Shift+Space`,
+say the sentence, release, and it is pasted into the terminal, editor or chat
+box that had focus. Transcription runs on whisper.cpp inside the app, so it
+works on a plane, on an offline machine, and with source code you cannot send
+to anyone.
+
+## What it does
+
+- **Push to talk** — global hotkey (remappable) with an always-on-top overlay
+  that shows the microphone level.
+- **Local transcription** — whisper.cpp via `whisper-rs`. Models download once
+  and live in the app-data directory.
+- **Pastes into the right window** — focus is captured before recording starts
+  and restored before the paste keystrokes.
+- **Snippets and vocabulary** — say "new line", get a newline; teach it that
+  your project is "Khan" not "can".
+- **AI cleanup (Pro)** — per-app modes that tidy the raw transcript using your
+  own API key or a local Ollama model.
+- **Agents can ask you (Pro)** — an MCP server exposes `ask_user_by_voice`, so
+  Claude Code or Cursor can put a question on your overlay and read your spoken
+  answer back.
+
+## Tiers
+
+|                             | Free                       | Pro ($59 once)              |
+| --------------------------- | -------------------------- | --------------------------- |
+| Minutes per day             | 30, resets at UTC midnight | Unlimited                   |
+| Models                      | Whisper base, small        | + large-v3-turbo            |
+| Modes, snippets, vocabulary | Snippets and vocabulary    | All, incl. AI cleanup modes |
+| `ask_user_by_voice` (MCP)   | —                          | Yes                         |
+
+The free tier is not a trial: there is no expiry and no account.
+
+## Install
+
+```bash
+bun install
+bun run tauri dev      # development
+bun run tauri build    # NSIS + MSI on Windows, DMG on macOS, AppImage/deb on Linux
+```
+
+On Windows the first build needs the [MSVC build tools](https://visualstudio.microsoft.com/downloads/#build-tools-for-visual-studio-2022),
+and `whisper-rs` compiles with CMake. `clang` must be on `LIBCLANG_PATH` for
+bindgen on every platform.
+
+## Using it
+
+1. Launch the app and hold the hotkey (default `Ctrl/Cmd+Shift+Space`).
+2. The overlay appears at the bottom of the screen; speak.
+3. Release. The transcript is pasted into the app you were in.
+
+Pick a model under **Settings → Model**. Whisper base is a ~140 MB download;
+large-v3-turbo is ~1.6 GB and noticeably better at code and proper nouns.
+
+## Settings worth knowing
+
+- **Hotkey** — any accelerator the global-shortcut plugin accepts. The new
+  combo is registered before the old one is dropped, so a bad combo never
+  leaves you without a working hotkey.
+- **Language** — leave on auto-detect unless you switch between languages often.
+- **Snippets / vocabulary** — matched on whole words, case-insensitively for
+  snippets. A trigger such as `c++` works too.
+
+## MCP integration (Pro)
+
+`agent-voice mcp` is a stdio MCP server that exposes one tool:
+
+```
+ask_user_by_voice({ prompt, timeout_ms? })
+```
+
+It talks to the running GUI over a loopback HTTP bridge (the port is published
+in `bridge.json` next to the app data), so the app must be running. Point any
+MCP client at the binary:
+
+```jsonc
+{
+  "mcpServers": {
+    "agent-voice": {
+      "command": "/path/to/agent-voice",
+      "args": ["mcp"],
+    },
+  },
+}
+```
+
+Without a Pro license the server starts but advertises no tools, so a
+misconfigured agent fails loudly instead of silently doing nothing.
+
+## Privacy
+
+Audio is captured only while the hotkey is held, written nowhere, and
+transcribed in-process. The only network traffic the app generates is:
+
+- downloading a model the first time you select it (from Hugging Face),
+- `POST /activate` when you enter a license code,
+- checking the update feed.
+
+Telemetry is not collected.
+
+## Development
+
+```bash
+bun run typecheck        # app + worker
+bun test                 # frontend helpers + worker (57 tests)
+bun run build            # frontend bundle
+cd src-tauri && cargo test   # Rust unit tests
+```
+
+The license backend is a Cloudflare Worker in [`worker/`](worker/README.md);
+the token format and key rotation are documented in
+[`docs/entitlement.md`](docs/entitlement.md).
+
+## License
+
+Source under the repository license. Whisper models are downloaded from
+Hugging Face under their own licenses (MIT for most, see the model cards).