Voice Command AI
GitHub

Version 0.1.0 · First public release

Your voice. Every agent. One command center.

Control every AI coding agent on your machine from your phone, with your voice.

A lightweight server on your machine owns every agent session. Your phone shows them as cards, and each command reaches the right session without stealing window focus.

View on GitHub User guide
  • MIT licensed
  • End-to-end encrypted
  • Runs on your machine, no VCA cloud

Illustration. The target session is shown before anything is sent.

Why VCA

Many agents. One keyboard. Something gives.

Parallel agent sessions bring sub-agents, compilers, test runners and helper windows onto one workstation. Three things break first.

  • Focus theft

    New windows keep stealing focus, so prompts and keystrokes land in the wrong terminal.

    With VCA

    Commands go straight to the target session through a virtual terminal. No simulated keystrokes.

  • Resource exhaustion

    Agents start builds and tests without knowing about each other, until CPU and memory are saturated and the machine stalls.

    With VCA

    A resource broker admits heavy jobs into a set number of slots and queues the rest.

  • Navigation overhead

    Remote-control apps make you drill into one project, back out and drill into the next, for every command.

    With VCA

    Every session sits on one deck of cards. Pick one, speak, send.

Features

What VCA does today

Everything below is in version 0.1.0, except the item marked “Coming soon”.

  • Session cards

    Every agent session on your machine as a live card: project, agent, status and alerts at a glance. Import AI CLI sessions that are already open, or start one in a folder you pick. Each card also counts the sub-agents its session runs.

  • Voice first

    Tap and speak. Live dictation streams as you speak; review the text, then send. A developer-vocabulary engine keeps technical terms intact.

  • No focus required

    Commands reach the target session through a virtual terminal or the agent’s own channel, never through simulated keystrokes.

  • Live view

    Follow any session as a readable chat or the raw terminal and keep the conversation going. Search the chat, read markdown tables, and scroll back through history that loads page by page.

  • Alerts, exact when you want

    A separate Alerts menu with a count badge lists every session waiting on a decision or input. Optional agent hooks, installed from Settings, make approval alerts exact. Answer with one tap.

  • Spoken alerts in your voice engine

    Hear alerts through your own OpenAI-compatible TTS engine, such as the reference server for MeloTTS or Supertonic. Pick a PC voice and everything spoken uses it, one item at a time, with a quick on/off. The device voice only fills in before the PC audio starts.

  • A to-do list from your agents

    When an agent needs you to sign in, approve or run something, the request lands in one list. Check it off and send the result back. Commands never run for you: admin tasks open an elevated shell only through your OS prompt.

  • Windows back after a restart

    After a reboot or a power cut, VCA reopens your AI CLI windows in the same place and the same conversation. Windows, macOS, Linux and tmux.

  • Remote shell, locked down

    When an agent is stuck, open a shell on your PC from the phone. Only admin-permission devices can, each shell needs Face ID or biometrics plus a fresh authenticator code, and it closes itself after a time limit.

  • Made for the phone

    Swipe through sessions in a carousel, attach photos, files or a quick sketch, and pick a light, dark or pure black theme. No PC yet? Try the demo with sample sessions, offline. In English, Korean, Japanese and Spanish.

  • Resource broker

    CPU, memory, disk and network per session. Heavy jobs are admitted, queued, throttled or paused by your policies.

  • Pairing and encryption

    One-time pairing links, TOTP two-factor sign-in, short-lived tokens, Noise end-to-end encryption, an audit log and a kill switch.

  • Your machines, your connection

    Pair your home, office and server machines with one app. Reach them over direct LAN, Tailscale, Cloudflare Tunnel, DNSPilot or your own provider. The server listens only on localhost until you pick one.

  • Bring your own engines

    Speech-to-text (a Whisper server, for example), a correction LLM and text-to-speech are yours to choose, local or in your own cloud account. VCA ships no speech servers and no model weights.

  • Screen layout

    Line up the terminal windows already open on your PC, monitor by monitor, in a grid of 1, 2, 4, 6, 8 or 10 cells or your own rows and columns. Auto fill, save layouts and apply them again, and undo the last move.

  • App lock

    The apps open locked and unlock with the device’s own sign-in: Windows Hello, Touch ID, Face ID, fingerprint or passcode. While locked, no session content shows and nothing can be sent. They lock again after 1 to 60 minutes without use.

  • Commonly misheard words

    Teach dictation the words it keeps getting wrong: each correct word with what gets written by mistake. The list lives on your PC, so it works on every connected device from your next dictation.

  • Coming soon

    Phone alerts while the app is closed

    Planned, not in this release. Today, phones show alerts while the VCA app is open, and desktops show system notifications in the background.

Command Center

A conductor for all your agents.

Turn on an optional conductor session and ask it “which sessions need me?” or “tell the api session to run the tests”. It answers from what VCA sees and acts through VCA’s MCP tools, only after you allow it.

  • Every message to another session waits for your Allow or Deny.
  • It never answers permission prompts. Those stay with you.
  • Actions are capped per hour, 20 by default.
  • A kill switch stops it and denies every pending request.
  • Every action is written to the audit log.

Runs on

  • Claude Code
  • Codex CLI
  • Gemini CLI
  • Built-in, with your own LLM

Off by default, and turned on only in the desktop app on the PC itself. How the conductor works

How it works

Your phone talks only to your machine.

After a one-time pairing, the app connects to the VCA server on your machine over an end-to-end encrypted Noise channel. The server owns the agent sessions. Nothing goes through a VCA cloud.

Watch the 2-minute walkthrough Video on YouTube (2:10). Opens in a new tab.
Architecture: the VCA app on your phone or desktop connects over an end-to-end encrypted channel to the VCA server on your machine, which manages agents such as Claude Code, Codex, Gemini CLI, Antigravity and Grok, and can use your own optional speech-to-text, LLM and text-to-speech engines. VCA app Phone · desktop Session cards Voice dictation Live view and alerts End-to-end encrypted Noise protocol LAN · Tailscale · Cloudflare DNSPilot · custom VCA server on your machine Session manager (PTY) Agent adapters Resource broker Speech pipeline Auth · audit · API Claude Code Codex Gemini CLI Antigravity Grok and more Agents on your machine Your speech engines (optional) STT · LLM · TTS, OpenAI-compatible Architecture: the VCA app on your phone or desktop connects over an end-to-end encrypted channel to the VCA server on your machine, which manages agents such as Claude Code, Codex, Gemini CLI, Antigravity and Grok, and can use your own optional speech-to-text, LLM and text-to-speech engines. VCA app Phone · desktop Session cards Voice dictation Live view and alerts End-to-end encrypted Noise protocol LAN · Tailscale · Cloudflare DNSPilot · custom VCA server on your machine Session manager (PTY) Agent adapters Resource broker Speech pipeline Auth · audit · API Agents on your machine Claude Code Codex Gemini CLI Antigravity Grok and more Your speech engines (optional) STT · LLM · TTS, OpenAI-compatible
Blue outlines mark agents with dedicated adapters. The others use the generic adapter.

  1. Run your agents under VCA

    Managed sessions accept input from the app without touching window focus.

    vca run -- claude

  2. Pair once

    In the desktop app, open Settings › Connection & security › Connect a phone and scan its QR code with the VCA app. One scan pairs the phone and, with DNSPilot, registers it for remote access. From a terminal:

    vca pair

  3. Pick a card and speak

    Choose the session, dictate, check the target and the text, then send.

Early measurements

Numbers from real machines.

Milestone M0 tested each building block on Windows, macOS, Linux and an iPhone in October 2026. A few results:

  • 7,199 / 7,199 requests answered in a 2‑hour soak over LAN and over Tailscale, with none lost
  • 1.3 s to reconnect after a server restart, with no re-pairing
  • ~0.3 s to transcribe about 10 s of speech on a GPU Whisper server
  • 5.5 ms median round trip on a LAN through the encrypted channel
  • Managed sessions verified on Windows and macOS
  • Real Claude Code hooks and Codex input
  • Admission control with real cargo builds
  • iPhone voice input to transcript

Measured on the first test builds in October 2026, before the 0.1.0 release. Read the full test report

Text-to-speech engines were compared separately: seven open-source engines, the same phrases in Korean, English, Japanese and Spanish. Read the TTS benchmark

Works with

Bring the agents you already use.

Agents

Any terminal agent can run as a managed session. Claude Code and Codex have dedicated adapters; the others use a generic adapter that reads prompts and waiting states from their output.

  • Claude Code
  • OpenAI Codex
  • Gemini CLI
  • Google Antigravity
  • Grok
  • Cursor CLI
  • Copilot CLI
  • and more

Dedicated adapter

Product names are trademarks of their respective owners.

Platforms

The server runs on your development machine; the app runs on your desktop and phone. Desktop installers are on GitHub Releases, and the phone apps are in review at the App Store and Google Play.

  • WindowsServer and desktop app · signed by MetaVu Co., Ltd.GitHub Releases
  • macOSServer and desktop app · signed and notarizedGitHub Releases
  • LinuxServer and desktop appComing soon
  • iOSMobile appApp Store · in review
  • AndroidMobile appGoogle Play · in review

Download

Get Voice Command AI 0.1.0.

Install the desktop app on the PC where your agents run. It includes the VCA server and the vca command, and updates itself from GitHub Releases; nothing is installed until you click. Then pair your phone with one QR code from Settings › Connection & security › Connect a phone.

  • Windows

    Installer (-setup.exe), signed by MetaVu Co., Ltd.

  • macOS

    Disk image (.dmg), signed and notarized.

  • Linux

    Coming in a later version. Until then, build it from source.

Windows (x64) macOS (Apple silicon) macOS (Intel)
App Store · in review Google Play · in review

All files and release notes on GitHub Releases

Or build it from source.

Install the server and the CLI from source, start the server, and run your first agent under VCA.

Requires Git and the Rust toolchain. On Windows, also the “Desktop development with C++” build tools.

$ git clone https://github.com/JinsongRoh/Voice-Command-AI.git
$ cd Voice-Command-AI
$ cargo install --path crates/vca-server   # vcad
$ cargo install --path crates/vca-cli      # vca
$ vca daemon
$ vca run -- claude

Open source

MIT licensed. Every feature, free.

The core is written in Rust and the apps are built with Tauri 2. Every feature works without paying anyone.

  • Read the code

    Server, CLI, shared UI and the desktop and mobile apps live in one repository under the MIT license.

    Browse the repository

  • Contribute

    Code, agent adapters, translations, or developer vocabulary for your language. Issues and pull requests are welcome.

    Open an issue

  • Supporting the project

    Every feature works without paying anyone. To help, consider DNSPilot for connectivity: it is built by the team behind VCA, and once its paid plans open, subscriptions will help fund VCA. It stays optional, on equal footing with Tailscale, Cloudflare Tunnel, LAN and custom providers. You can also contribute, or star the repository and share it.

    Star the repository

VCA is built on the work of many open-source projects. Thank you to everyone behind them.