STP — Hands-Free Voice Assistant for Claude Code and Antigravity

Wake it with a phrase. Dictate. Send by voice. STP works the way "hey Siri" works — except instead of opening apps or setting timers, it talks to your AI coding assistant for you. Works with Antigravity or Claude Code. You pick the target, you pick the chat.

100% local. Speech is transcribed on your machine with Vosk (offline, WebAssembly). There is no API key, no cloud transcription, and nothing you say ever leaves your computer. The only network access is a one-time ~40 MB speech-model download on first run.

Internally the package is named VTP (Voice to Prompt). It's the same project — STP is the user-facing name, VTP is the codename / repo / package id.

Watch the Demo

Click to open on YouTube

How It Works

The full loop — start, dictate, send, repeat — happens by voice. No mouse. No keyboard. No focus required.

Say your wake phrase and STP starts listening. The default is "hey antigravity" — change it in Settings. Then dictate naturally and end with something like "send the prompt." STP routes it into either Antigravity or Claude Code, whichever you've selected as the target.

Everything runs on-device:

• Speech-to-text — Vosk runs as WebAssembly inside the extension's webview. Your microphone audio is transcribed locally and never uploaded.
• Filler cleanup + intent — pure local regex. Instant, offline.
• Prompt "enhance" (optional) — if you have Ollama running locally, STP uses it to rewrite dictation into a spec-style prompt. If Ollama isn't installed, "enhance" gracefully falls back to a local cleanup pass. No key either way.

Two listening modes

Switch in the Settings menu inside the panel.

Pick your target — Antigravity or Claude Code

A button in the panel header (→ AG / → CC) toggles which AI chat receives your prompts. The voice flow doesn't change — only where the prompt lands.

When the target is Claude Code, the context card below the button becomes a chat lock. You can have several Claude Code conversations open and only the one you locked will receive the prompt.

Send while you're tabbed out

Injection doesn't need focus. Start a prompt by voice, alt-tab to a browser, say "send the prompt" — it lands. The chat is in your IDE, not your active window.

What You Can Say

Auto-clean before send

Filler words ("um", "uh") and immediate repeats are stripped locally before every send. If the buffer is genuinely noisy (self-corrections, "actually", "I mean") and you have Ollama installed, STP runs a small local cleanup pass. Otherwise it's regex-only — zero setup, zero network.

Setup

1. Install the extension

Install from Open VSX / the marketplace, or code --install-extension vtp-voice-prompt.vsix.

2. First run downloads the speech model

The first time you open the STP panel it downloads a Vosk English model to the extension's local storage. You'll see a progress banner. After that it works fully offline — the model is cached and reused.

It is a 130 MB download and STP cannot hear you until it finishes. That is the accurate model, and on a slow connection it is a long first run. If you would rather be talking in a minute, set vtp.voskModelUrl to the small one before you open the panel:

https://ccoreilly.github.io/vosk-browser/models/vosk-model-small-en-us-0.15.tar.gz

40 MB, noticeably less accurate on unusual words, and swappable later — changing the setting and running VTP: Re-download Speech Model fetches the other one.

Another language, or a different size? Point vtp.voskModelUrl at any vosk-browser-compatible .tar.gz, or an alphacephei .zip — those are converted on first run.

3. (Optional) Install Ollama for smart "enhance"

If you want the enhance this prompt command to do context-aware rewriting, install Ollama and pull a model:

ollama pull llama3.2

STP auto-detects Ollama on 127.0.0.1:11434. No key, no config — nothing leaves your machine. Without Ollama, "enhance" just runs a local cleanup pass. Change the model with vtp.enhancementModel.

4. Dictate

Open the STP panel from the Activity Bar. Click the microphone (or use Ctrl/Cmd+Shift+Space) and speak. The transcript updates live as you talk.

Claude Code Integration

Switch the target to Claude Code with the → AG / → CC header button. The first time, run "VTP: (Re-)apply Claude Code Patch" from the command palette once (it injects a tiny runtime hook so prompts can be delivered into the Claude Code webview) and reload the window.

• Lock a conversation — "VTP: Lock Claude Conversation (Pick Target Chat)" so prompts only land in that chat.
• Check it is working — "VTP: Check Claude Code Integration" says whether prompts can actually be delivered right now, and if not, which part is wrong.
• Restore — "VTP: Restore Claude Code (Unpatch)" rolls back to the pre-VTP state.

Because transcription is now fully local and fast, Claude Code mode no longer needs any third-party streaming service.

When Claude Code updates

Claude Code is a minified bundle, so the hook is attached by matching patterns in it — and an update can move what those patterns match. VTP watches for a new version, patches it within seconds of it landing, and offers you the reload. It writes what it found to %APPDATA%\VTP\health.json:

{ "claudeCode": { "version": "2.1.233" },
  "patches":  { "extJs_commands": "applied", "wvJs_handler": "applied" },
  "commands": { "inject": true },
  "working":  true }

working: false names the part that is wrong. Two failures look identical from the outside and want opposite things done, so they are reported separately: an anchor that no longer matches needs the pattern updating, while patched files with the commands missing just means this window started before the patch and a reload fixes it.

Sending a prompt from another program

Anything running as you can hand a prompt to an open Claude conversation by dropping a file in %APPDATA%\VTP\inbox:

{ "text": "what changed in the last commit?",
  "submit": false,
  "targetTitle": "vendor email" }

submit sends it; left out, it sits in the composer for you to read first. targetTitle picks a conversation by part of its name — leave it out and it goes to whichever Claude chats are open. A <name>.done.json appears beside it saying what happened, including which conversations received it, so the caller can tell delivery from silence. A name matching nothing is refused with a list of the ones that are open, rather than reported as sent.

Settings

Commands

Privacy

Requirements

• VS Code / Antigravity ^1.85.0
• A microphone
• FFmpeg on your PATH — used to capture the mic locally (the webview can't access the microphone directly). The panel prompts you to install it if it's missing.
• ~40 MB free space for the cached speech model
• (Optional) Ollama for smart prompt enhancement

No cloud account, no API key — FFmpeg and the speech model are the only local dependencies.

License

MIT