PRMPTR

Real-time AI conversation wingman. PRMPTR listens to your microphone ("You") and system audio ("Them"), transcribes both on your device, and streams context-aware suggestions — answers, comebacks, facts — in the personality you pick.

Built with Tauri 2 · Next.js 16 (static export) · React 19 · Rust.

Highlights

• On-device speech recognition — transcribe.cpp runs GGUF models locally; nothing leaves your machine.
  • Parakeet Unified (English) by default: ~1.6% WER, punctuated, >15× real time on an ordinary desktop CPU.
  • Also: Parakeet v3 (25 European languages), Whisper Large v3 Turbo (99 languages), Moonshine Small and Parakeet 110M for low-power machines.
  • Silero VAD v6.2 gates the models so silence and room noise never reach them (no "Thank you." hallucinations).
  • Models download on demand from pinned revisions and are SHA-256 verified.
  • Personal glossary — names and jargon bias recognition, and known mishearings are corrected (Saoirse = Sayers).
  • Speaker names — rename "Speaker 2" to "Sarah" in the live feed; the AI sees who said what.
  • Optional GPU builds via Vulkan (NVIDIA/AMD/Intel) or CUDA.
• Deepgram as the one cloud transcription option.
• Your AI apps — use the subscriptions you already have (Claude Code, Codex, OpenCode, Gemini CLI, Ollama, LM Studio…) through OriginLM. Turns run through the installed client; PRMPTR never reads or copies your logins.
• API providers — Anthropic (Claude Haiku/Sonnet/Opus 5.5), OpenAI (GPT-6.x), Groq, Cerebras, OpenCode Zen, or a local LM Studio server. Live model lists come from each provider; Groq ⇄ Cerebras fail over on rate limits.
• Always-on-top overlay (opt-in) — suggestions stream into a small floating window that's hidden from screen sharing; Ctrl+Shift+H show/hide, Ctrl+Shift+C click-through.
• 10 personalities, global shortcuts, optional screenshot context for vision models, local or HTTP text-to-speech, and session history stored locally.

Getting started

Prerequisites: Node.js 22+, the Rust toolchain (MSVC on Windows), CMake, and the Tauri prerequisites. On Linux, system audio needs a PulseAudio or PipeWire monitor source and parec (package pulseaudio-utils on Debian/Ubuntu, or pipewire-pulse).

npm install
npm run tauri dev          # desktop app with hot reload (http://127.0.0.1:43110)
npm run tauri build        # production installers (NSIS/MSI, .deb/.rpm)
npm run tauri:dev:gpu      # dev build with the Vulkan GPU backend (needs the Vulkan SDK)

First run: Settings → Capture & Voice → download Parakeet Unified (731 MB, one time) → pick your microphone and output device → Start. Then pick a model under Providers & Models (or let Your AI apps find the ones you already have) and press Analyze (Ctrl+Shift+Space).

Testing

Command	What it covers
npm run typecheck / npm run lint	TypeScript 6 strict, ESLint 10
npm test	Vitest unit tests (prompting, model migration, providers, …)
npm run test:e2e	Playwright against the production static export with an in-page fake of the Rust backend and intercepted provider APIs
npm run test:rust	Rust unit tests (VAD segmenter, model catalog, OriginLM mapping, …)
PRMPTR_TEST_MODELS=<dir> cargo test --test real_audio	Real audio through Silero VAD v6.2 + Parakeet
npm run test:all	Everything above except the real-audio test

Troubleshooting

• Fedora 44, KDE Wayland, NVIDIA: the window flashes and closes. WebKitGTK 2.54.0/2.54.1 has an upstream DMA-BUF regression (WebKit bug 324551) fixed in 2.54.2. PRMPTR detects that combination before creating the WebView and offers compatibility mode; updating is the real fix:

  sudo dnf upgrade webkit2gtk4.1 javascriptcoregtk4.1

  PRMPTR_WEBKIT_COMPAT=0 skips the check; =1 forces compatibility mode.
• No "Them" audio on Linux. Install parec (sudo dnf install pulseaudio-utils / sudo apt install pulseaudio-utils); it records the default sink's monitor source.

Privacy

• Transcription is on-device unless you choose Deepgram.
• Transcripts and history stay in local IndexedDB; nothing is synced.
• API keys live in a Tauri-managed store file, never in localStorage.
• Your AI apps never exposes subscription credentials to PRMPTR.
• The webview can only reach the provider APIs and loopback addresses allowed in src-tauri/capabilities/default.json.

See docs/architecture.md for the full design.

Project structure

app/            Next.js routes (static export — no API routes)
components/     React UI (dashboard, analysis, settings, model picker)
lib/            Stores, LLM client/providers, Tauri bridge, prompt builder
src-tauri/      Rust: audio capture, transcribe.cpp engine + VAD, Deepgram,
                OriginLM bridge, screenshots, TTS
e2e/            Playwright suite and the fake Rust backend