A native Windows overlay that listens to your meetings, transcribes speech in real time, and answers technical questions through an LLM — cloud or fully local — in small floating windows ("tiles") beside your meeting window.
Built in pure Rust + Slint (skia renderer, no browser engine, no Node). Transparent, always-on-top, with Windows capture exclusion available through an explicit stealth toggle.
Windows 10/11 only. Single user, no telemetry, no code signing. Interface languages: English and Russian (switchable at runtime).
Latest published build: v0.38.0.
The master branch may contain unreleased work.
 |
 |
 |
 |
 |
 |
 |
 |
- Cloud or local AI. Use a cloud LLM (Claude via an OpenAI-compatible bridge) or install llama.cpp with an in-app choice of Gemma 4 E4B / 12B / 26B-A4B profiles (CUDA or CPU). Combine it with local STT for an offline meeting pipeline. The standalone PowerShell script installs the lightweight 4B profile.
- Speech-to-text. Cloud (Groq Whisper), local whisper.cpp server, or GigaAM-v3 running in-process (CPU or DirectML GPU). Groq and whisper.cpp support mixed Russian + English; GigaAM is the Russian specialist.
- Transparent overlay bar. Always-on-top HUD showing session status, live transcript, mic/system-audio toggles, timer, and action chips.
- AI tiles. Markdown-rendered answers with pin, maximize, copy, follow-up conversation, and adjustable opacity.
- Auto-tiles. Question/keyword detector in the transcript spawns answer tiles automatically (configurable; can skip your own mic input).
- Meeting summary. One-click structured summary of the full transcript. Long sessions are chunked and processed via map-reduce without truncation.
- Audio recording. Mic and system audio recorded to separate WAV files with a configurable retention policy. Offline re-transcription and summary rebuild from the archive.
- Session archive (F7). Full-text search (SQLite FTS5) across past sessions — transcripts and AI answers. Includes a built-in audio player with click-to-seek and per-line highlighting.
- Personal memory. Your facts, terms, and names (internal projects, abbreviations) injected into AI answers and summaries. Candidates extracted from sessions with manual review.
- Vision AI (F8). Select a screen region for analysis or description. Shift+F8 translates the captured text. Ctrl+F8 runs OCR for the read-aloud pipeline.
- Read-aloud (TTS). Neural text-to-speech via a separate sidecar process (sherpa-onnx). Shift+Alt+1 reads selected text, Shift+Alt+2 reads an OCR region, Shift+Alt+3 pauses/resumes.
- Speaker diarization. Offline speaker segmentation (pyannote + WeSpeaker) through the TTS sidecar, available from the archive.
- Knowledge base palette (F4). A large built-in library (glossary, commands, patterns). Search and open as a tile — no AI call, zero cost.
- Stealth mode. Windows request capture exclusion through
WDA_EXCLUDEFROMCAPTURE, used by modern Windows capture APIs. Capture software can vary, so verify it with your own meeting setup before relying on it. - Auto-update. Checks GitHub Releases, downloads the installer, verifies its SHA-256 digest against the release metadata, and launches it. Verification is fail-closed, and downloads are restricted to GitHub hosts.
- Hermes plugin. Optional two-way integration with a local Hermes agent instance — install the plugin from Settings.
- Context window control. Auto / 8K / 16K / 32K / 64K / 96K presets for the managed local llama.cpp server, with hardware-aware memory estimates.
- Download
suflyor-slint-setup.exefrom GitHub Releases. - Run it. The binary is unsigned, so SmartScreen may warn: More info → Run anyway.
- Installs to
%LOCALAPPDATA%\suflyor-slint\— no admin rights required. A Start menu shortcut is created. - Launch the app. A seven-step wizard walks through cloud/local mode, AI, speech recognition, microphone, system audio, and overlay preferences.
Configuration is stored in %APPDATA%\suflyor\config.json.
From the app: Settings → AI bridge → Install / complete local AI. This downloads llama.cpp, whisper.cpp, and models, detects your GPU (CUDA), starts the servers, and writes the settings.
Or run the standalone script:
powershell -ExecutionPolicy Bypass -File scripts\setup-local-ai.ps1What gets installed (into %USERPROFILE%\suflyor-local-ai):
| Component | Role | Endpoint |
|---|---|---|
| llama.cpp + Gemma 4B | LLM | http://127.0.0.1:8080/v1 |
| whisper.cpp + Whisper large-v3-turbo | STT (mixed RU+EN) | http://127.0.0.1:8081/v1 |
| GigaAM-v3 | STT (best Russian) | in-process, no server |
Script flags: -Cpu (force CPU), -NoLaunch (download only),
-SkipLlama / -SkipWhisper / -SkipGigaam. Re-running resumes partial
downloads and skips completed components.
- Complete the first-launch wizard. Each hardware/service step has its own test, and Settings → Diagnostics → Check all is the single place to confirm the complete stack later.
- Treat the long top bar as the control panel: microphone and speaker chips choose what Suflyor hears; Start begins a session; + tile opens a manual answer window.
- Speak normally. The second line of the bar shows the latest transcript and whether it came from your microphone or system audio. Automatic answers appear as tiles when question detection is enabled.
- Press F1 for the in-app help, or F7 to search past sessions. Every shortcut is also listed in Settings → Hotkeys; Settings → Diagnostics shows whether each one registered successfully.
| Key | Action |
|---|---|
| F1 | Toggle help window |
| F3 | Re-ask the last question with fresh context |
| F4 | Toggle knowledge base palette |
| F6 | Manual tile from the last transcript line |
| F7 | Toggle session archive |
| F8 | Screenshot → Vision AI (select a region; Esc cancels) |
| Shift+F8 | Screenshot → translate captured text |
| Ctrl+F8 | Screenshot → OCR / read-aloud |
| F9 | Ask the AI now (streaming answer) |
| Shift+F9 | Ask with cloud escalation (deeper reasoning, one-shot) |
| Shift+Alt+1 | Read selected text aloud (clipboard) |
| Shift+Alt+2 | Read an OCR region aloud |
| Shift+Alt+3 | Pause / resume read-aloud |
Push-to-talk: hold ask for your microphone or grab for system audio.
Three standalone crates (no root workspace):
slint-experiment/ overlay-host binary — Slint UI (ui/*.slint compiled
in via build.rs), Win32 transparency/stealth, hotkey
dispatch, tile lifecycle, settings orchestration
overlay-backend/ shared no-UI library — audio (WASAPI loopback + mic),
STT dispatch (Groq / whisper.cpp / GigaAM), AI client
(OpenAI-compatible, streaming, retry, cost tracking),
local AI installer, vision, OCR, TTS/diarization control,
journal, KB, memory, config, auto-update
suflyor-tts/ neural read-aloud + diarization worker — links
sherpa-onnx only; must stay a separate process (two
static onnxruntimes in one binary crash)
Data flow: WASAPI capture → STT → transcript ring buffer → question/keyword detector → prompt build (with KB + memory injection) → AI (local or cloud) → tile spawn on the UI thread.
Overlay windows use Win32 WS_EX_LAYERED / DWM blur-behind /
SetWindowDisplayAffinity for transparency, always-on-top, and stealth.
See docs/architecture.md for the full developer
overview and CONTRIBUTING.md for the build setup.
- No telemetry. Nothing is collected or reported.
- Local-first option. When AI and STT are both local, meeting audio, transcripts, and prompts are processed on your machine.
- Secrets stay local. API keys are stored in
%APPDATA%\suflyor\config.jsonand sent only to the services you configure. The app avoids logging secret values, and UI errors hide URLs and hosts. - Update downloads are verified. The auto-updater only accepts GitHub-hosted assets and verifies the installer's SHA-256 digest against GitHub release metadata before execution. The installer is not Authenticode-signed.
- Stealth is explicit. Screen-capture exclusion is off by default and toggled in Settings.
- Windows only. WASAPI for audio capture, Win32 for window management — there is no macOS or Linux support today.
- Roadmap. Native Linux and macOS versions are planned, but no builds or delivery dates are promised yet.
- No code signing. SmartScreen will warn on first launch. This is a deliberate choice for a single-user tool.
- Single user. No multi-user profiles, no concurrent sessions.
- GPU-dependent local AI quality. The 26B Gemma profile needs a GPU with sufficient VRAM; CPU fallback uses smaller models.
- Stealth vs. interaction. Stealthed windows force an arrow cursor to avoid leaking the overlay's custom cursor into screen captures.
- Capture exclusion is not a universal guarantee. Test stealth with the exact meeting/recording software and Windows version you use.
Prerequisites: Windows 10/11, Rust (stable-msvc toolchain), Visual Studio Build Tools 2022 with the C++ workload.
# Dev build + run
cargo run --bin overlay-host --manifest-path slint-experiment\Cargo.toml
# Release build + NSIS installer
powershell -ExecutionPolicy Bypass -File scripts\build-slint-release.ps1 -Installer
# → slint-experiment\target\release\bundle\suflyor-slint-setup.exe
# Full CI gate (fmt + clippy + tests for all 3 crates + i18n guard)
powershell -NoProfile -ExecutionPolicy Bypass -File scripts\ci.ps1Suflyor is a 100% vibe-coded product. The product owner has not manually read or written a single line of its source code. Their role has been to choose the stack, define the product direction, and make technical decisions. The implementation was produced with Codex, Claude Code, and Qwen Code.