Vesper

Desktop meeting notes that stay on your machine. Vesper records microphone and system audio as two separate channels, transcribes them live, and writes a summary — with no account, no telemetry, and nothing sent anywhere unless you ask for it.

No bot joins your call. Nothing is uploaded to be processed. The transcript, the summary and the audio are files on your disk.

Left: the transcript, cut at pauses rather than on a clock. Right: the summary the local model wrote from it.

• Record, pause, resume, stop — with a floating card that stays on top while you work in another window
• Dual-channel capture: microphone is Me, system audio is Others
• Live transcript, cut at pauses rather than on a clock, so a bubble is a sentence
• Import an audio file, or retranscribe one you already have
• Summary templates, action items, key points, and a chat over the meeting
• Notes you type during the meeting, folded into a corner until you need them
• Search across every meeting; export to Markdown, PDF or DOCX
• System tray, a global hotkey you can change, and update checks
• Quiet-room guard: three minutes without speech asks whether you are still there, five more without an answer stops the recording

This is the complete list. Four situations, and one of them happens without you asking:

No analytics, no crash reporting, no "anonymous usage statistics". No fifth destination.

Local by default. Out of the box, transcription runs on whisper.cpp and summaries on llama.cpp, both on your machine. OpenRouter is a choice you make in Settings, never a fallback the app takes by itself.

Offline mode is the enforcement, because "cloud optional" is only a promise if something checks. With the switch on, cloud transcription, cloud summaries, model downloads and the update check are all refused before a socket opens, and each refusal names the thing that did not happen rather than reporting a connection error you would go looking for.

It deliberately does not block an OpenAI-compatible endpoint you configured yourself. The reasoning is that a model on your own machine is not egress — but if you pointed that endpoint at another machine on your network, offline mode will not stop it, and you should know that rather than discover it.

API keys go to the operating system's keychain — Credential Manager on Windows, Keychain on macOS, the Secret Service on Linux. If the keychain refuses a key you just typed, saving fails and tells you, rather than quietly writing the key somewhere less safe. A key from an older version that has not migrated yet stays in the local database until a keychain write succeeds.

Installers are produced for Windows, macOS and Linux. They are not equally exercised, and pretending otherwise would be dishonest:

Day-to-day use, and every manual check of the interface, happens on Windows 11. CI compiles the project and runs the Rust test suite on Linux and Windows, so a Linux build that fails to compile or breaks a test is caught — but nobody has watched the app record a meeting there. macOS is built by the release workflow and is otherwise unverified: it has no CI job, and the Apple bundle has never been launched by the maintainer.

Bug reports from macOS and Linux are genuinely useful, and are the fastest way for that to change.

Optional, off by default, and a build-time choice — --features gpu-vulkan. The Vulkan SDK is needed to compile it and never to run it: vulkan-1 ships with every GPU driver, so an installed build asks the user for nothing. Without it, everything runs on the CPU, slower and identically.

• Desktop: Tauri 2 + Rust
• UI: React 19, Vite, TypeScript, Tailwind CSS 4, Framer Motion
• Speech: whisper.cpp via whisper-rs
• Language models: llama.cpp via llama-cpp-2
• Data: SQLite (rusqlite), and audio files next to it

Prerequisites: Node.js 20+ (CI builds on 22), Rust, and the platform WebView dependencies (Tauri prerequisites).

On Debian or Ubuntu, audio capture and the native build need five packages that are not on Tauri's prerequisites page — CI installs exactly these:

sudo apt-get install -y libasound2-dev libpipewire-0.3-dev libclang-dev cmake pkg-config

Native ML build (whisper.cpp + llama.cpp) needs LLVM for libclang. On Windows the build script also wants LLVM's nm and objcopy, which means the whole directory on PATH rather than one variable per tool — it asks for the next tool only after finding the previous one:

winget install -e --id LLVM.LLVM
$env:LIBCLANG_PATH = "C:\Program Files\LLVM\bin"
$env:Path = "C:\Program Files\LLVM\bin;" + $env:Path

npm install
npm run tauri dev

Whisper and GGUF weights are downloaded from Settings, from a fixed internal catalogue — the front end cannot ask for an arbitrary URL — and every file is checked against a known hash before it is loaded. Until a model is installed the local engines have nothing to run: there is no fake transcription to fill the gap.

npm run typecheck              # tsc --noEmit
npm run build                  # typecheck + production UI build
cd src-tauri && cargo test     # Rust unit tests
npm run tauri build            # desktop package

Building with Vulkan on Windows needs the Ninja generator and a short target directory — the nested vulkan-shaders-gen project trips MSBuild's parallel scheduling, and its path outgrows the 260-character limit that tracker.exe still enforces:

CMAKE_GENERATOR=Ninja CARGO_TARGET_DIR=/c/vt cargo build --features gpu-vulkan

Release builds check https://github.com/Yefclub/Vesper/releases/latest/download/latest.json. The signing public key lives in src-tauri/tauri.conf.json under plugins.updater. The check runs at launch and is governed by Check for updates at launch in Settings, on by default; turning it off also stops the installer being fetched ahead of the click.

MIT — see LICENSE.