Voice-controlled AI agent overlay for your entire computer.
Built with Tauri 2, Rust, and Svelte 5.
Website • Discord • Features • Quick Start • Providers • Changelog • License
Alpha Release — Voice Mirror is in active development. Core features work and are tested across all platforms, but expect rough edges. Bug reports and feedback welcome via GitHub Issues or Discord.
Claude Code / OpenCode = Terminal + MCP Tools
Voice Mirror = Eyes + Ears + Voice
Combined = Full AI agent for your entire computer
| Existing Product | What's Missing |
|---|---|
| Siri / Cortana / Alexa | No real capabilities, can't "see" your screen |
| GitHub Copilot | Code only, no voice, IDE-locked |
| Claude Desktop | Not an overlay, no wake word, limited platform |
| ChatGPT Desktop | Just a chat window |
Voice Mirror is an always-on overlay that listens, sees your screen, executes commands, and speaks back — across any AI provider.
- Tauri 2 + Rust native backend — no Electron, no Node.js runtime in the app. The entire backend is Rust: window management, voice pipeline, AI provider orchestration, config, IPC. Minimal memory footprint and instant startup.
- Rust-native voice pipeline — STT (Whisper ONNX), TTS (Kokoro ONNX / Edge TTS), and VAD run directly in the Tauri process via rodio. No child processes, no FFI bridges, no external binaries.
- Svelte 5 frontend — reactive UI with Vite HMR. Runes-based stores (
$state,$derived,$effect) for real-time state management. - Claude Code / OpenCode as PTY-backed brains — runs Claude Code or OpenCode inside the app as a real terminal. Full MCP tool access with 75+ models via OpenCode, zero extra API cost beyond the CLI itself.
- Native Rust MCP server —
voice-mirror-mcpis a compiled Rust binary using stdio JSON-RPC. Named pipe server for fast MCP-to-app communication with file-based inbox fallback. 58 tools across 10 dynamically-loaded groups. - Real browser automation — not "search and summarize." Actual CDP-level click, type, navigate, screenshot, and DOM snapshot.
- Unified agent loop — wake word + screen capture + terminal execution + browser control + persistent memory, all wired together. Most tools pick one; this closes the full Hear > See > Think > Act > Speak > Persist loop.
| Mode | Trigger | Use Case |
|---|---|---|
| Wake Word | "Hey Claude" | Hands-free, on-demand |
| Push to Talk | Mouse button / keyboard | Manual control |
| Dictation | Configurable hotkey | Voice-to-text input anywhere |
Capture your screen at any time — the AI sees what you see and can analyze errors, UI, or anything on screen. Multi-monitor support. Images are sent directly to vision-capable LLMs (Ollama, OpenAI, etc.).
Full chat interface with message history, Slack-style grouped messages, markdown rendering, image paste/drop, and persistent chat sessions. AI activity status bar shows real-time provider state.
Claude Code or OpenCode runs inside Voice Mirror with full MCP tool access. Execute commands, manage files, and automate workflows — all by voice. OpenCode unlocks 75+ models (GPT, Gemini, Kimi, local Ollama, and more) through a single integration.
| Group | Tools | Capabilities |
|---|---|---|
| core | 4 | Voice I/O, presence tracking |
| memory | 6 | Semantic search, 3-tier persistent memory |
| browser | 16 | Full CDP automation — navigate, click, type, screenshot, cookies, storage |
| n8n | 22 | Workflow CRUD, executions, credentials, tags |
| voice-clone | 3 | Clone any voice from a 3-second audio sample |
| screen | 1 | Desktop screenshot capture |
| diagnostic | 1 | Pipeline message tracing |
| meta | 3 | Dynamic tool loading/unloading |
| facades | 3 | Single-tool wrappers for memory, browser, and n8n (voice-mode efficiency) |
| Engine | Speed | Notes |
|---|---|---|
| Whisper (default) | Fast | Whisper ONNX, runs natively in the Rust backend |
| Engine | Speed | Voice Cloning | Voices |
|---|---|---|---|
| Kokoro (default) | Fast, CPU | No | 10 built-in |
| Edge TTS | Fast, online | No | 400+ Microsoft voices |
10 built-in providers — 75+ models accessible through OpenCode. Local servers are auto-detected.
| Provider | Type | Key Features |
|---|---|---|
| Claude Code | CLI agent | Anthropic's gold-standard CLI. MCP tools, vision, full terminal, CLAUDE.md ecosystem |
| OpenCode | CLI agent | Universal gateway to 75+ models (GPT, Gemini, Kimi, Grok, Mistral, and more). Full MCP tool support |
| Codex | CLI agent | OpenAI's CLI agent |
| Gemini CLI | CLI agent | Google's CLI agent |
| Kimi CLI | CLI agent | Moonshot's CLI agent |
| Ollama | Local API | Auto-detect, vision (llava), run models on your own hardware |
| LM Studio | Local API | Auto-detect, GUI-based local LLM runner |
| Jan | Local API | Auto-detect, open source local AI |
| OpenAI | Cloud API | GPT models via API |
| Groq | Cloud API | Ultra-fast inference |
Grab the latest installer from GitHub Releases:
| Platform | Download |
|---|---|
| Windows | Voice-Mirror-Setup-x.x.x.exe (NSIS installer) |
| macOS | Voice-Mirror-x.x.x.dmg |
| Linux | Voice-Mirror-x.x.x.AppImage |
Everything is bundled — no Node.js, Rust, or build tools required. The app checks for updates automatically and notifies you when a new version is available.
For contributors or running from source:
git clone https://github.com/contextmirror/voice-mirror-electron.git
cd voice-mirror-electron
# Install JS dependencies (for tests and tooling)
npm install
# Run the Tauri app with hot-reload
cd tauri
cargo tauri devDev requirements: Node.js 22+, Rust toolchain, LLVM/libclang, CMake
Verify builds:
cd tauri && npx vite build # Check Svelte frontend
cd tauri/src-tauri && cargo check # Check Rust backend
cd tauri/src-tauri && cargo test # Run Rust tests (167)
npm test # Run JS tests (1070+)- Claude Code CLI (for Claude provider) or OpenCode (for 75+ models)
- Codex CLI, Gemini CLI, or Kimi CLI (for additional CLI providers)
- ffmpeg (for voice cloning)
- CUDA (optional — GPU acceleration for supported models)
| Shortcut | Action |
|---|---|
Ctrl+Shift+V |
Toggle panel expand/collapse |
Ctrl+Shift+M |
Toggle performance monitor |
| Drag orb | Move orb position |
Config is stored per-platform:
| Platform | Path |
|---|---|
| Linux | ~/.config/voice-mirror/config.json |
| macOS | ~/Library/Application Support/voice-mirror/config.json |
| Windows | %APPDATA%\voice-mirror\config.json |
Settings are accessible from the in-app Settings page — AI provider, voice, activation mode, audio devices, appearance, and more.
voice-mirror/
├── tauri/ # Tauri 2 application
│ ├── src-tauri/ # Rust backend
│ │ ├── src/
│ │ │ ├── commands/ # Tauri commands (config, window, voice, ai, chat, tools, shortcuts)
│ │ │ ├── config/ # Config management + schema
│ │ │ ├── providers/ # AI providers (CLI PTY + API HTTP)
│ │ │ ├── voice/ # Voice pipeline (STT, TTS, VAD) — runs natively
│ │ │ ├── mcp/ # Built-in MCP server (Rust binary)
│ │ │ └── ipc/ # Named pipe server
│ │ ├── Cargo.toml # Rust dependencies
│ │ └── tauri.conf.json # Tauri config
│ └── src/ # Svelte 5 frontend
│ ├── components/ # Chat, settings, sidebar, overlay, terminal, shared
│ ├── lib/stores/ # Reactive stores (.svelte.js with $state/$derived/$effect)
│ ├── lib/ # Utilities (api.js, markdown.js, utils.js)
│ └── styles/ # CSS (tokens.css, settings.css)
├── electron/ # Legacy Electron app (still in repo, not the active target)
├── mcp-server/ # Legacy Node.js MCP server
├── test/ # Test suites
│ ├── tauri/ # Tauri-era JS tests (1070+)
│ ├── unit/ # Legacy unit tests
│ └── integration/ # Legacy integration tests
├── chrome-extension/ # Browser relay extension (MV3)
├── cli/ # Setup wizard + CLI
└── assets/ # Icons
Developer:
"Hey Claude, what's this error on my screen?" -> captures screen, analyzes code, suggests fix
"Fix it" -> executes commands in terminal
Desktop:
"What app is using all my memory?" -> checks processes, reports findings via voice
Voice Cloning:
"Clone David Attenborough's voice from this clip" -> downloads audio, creates clone — all responses now in that voice
Automation:
"Create an n8n workflow that emails me when my server goes down" -> builds workflow via MCP tools
| Platform | Status | Notes |
|---|---|---|
| Linux | Primary | X11/Wayland, AppImage target |
| Windows | Supported | NSIS installer |
| macOS | Supported | DMG target |
Voice Mirror runs its own isolated browser environment. It does not attach to or control your existing browser.
- Embedded Chromium — browser automation operates in a controlled instance launched by the app, not your system browser. No access to your existing sessions, cookies, or logged-in accounts unless you explicitly log in inside Voice Mirror.
- Tool-mediated actions — every browser action flows through
LLM -> tool schema -> browser controller -> Chromium. Actions are enumerable, reviewable, and loggable. No arbitrary JS injection. - Explicit screen capture — screenshots are triggered by tool calls or user request, not captured passively. The orb visually indicates capture state.
- MCP tool gating — tool groups load on demand. The LLM can only access tools that have been explicitly loaded for the session.
| Variable | Provider |
|---|---|
ANTHROPIC_API_KEY |
Claude Code |
Cloud provider API keys (OpenAI, Google, xAI, etc.) are configured through OpenCode directly — see OpenCode docs for setup.
npm test # JS tests (1070+ across test/tauri/)
cd tauri/src-tauri && cargo test # Rust tests (167)1070+ JavaScript tests (node:test + node:assert/strict) covering config safety, API key detection, provider detection, settings, startup behavior, cross-platform paths, terminal rendering, MCP memory, browser automation, IPC validation, structured logging, path safety, and more. 167 Rust tests covering commands, config, providers, voice pipeline, and MCP server.
Built with Tauri, Rust, Svelte, and a lot of voice commands.