Mac
- Mac: Apple Silicon — M1 or later.
- Operating system: macOS 15.0 or later.
- Install: download the DMG, or run
brew install --cask stevyhacker/tap/lokalbot. - Disk: room for the app, recordings, and selected models.
- Internet: required for the initial app and model downloads; optional for the built-in workflow after setup.
- Library:
~/Library/Application Support/me.dotenv.LokalBot.
Intel Macs are not supported. The Mac app is built around Apple Silicon acceleration, Core Audio process taps, the Neural Engine, and Metal. Windows and Linux have their own native app, built separately, with the same pages and the same library format.
How much memory does a Mac need?
The Mac app runs across the supported Apple Silicon range, but local language-model choice changes the practical memory requirement. Compact built-in models start around 0.53–1.28 GB. Several balanced summary and cotyping models recommend 16 GB of unified memory. The largest long-meeting models in the catalog are roughly 16.8–17.73 GB on disk and target Macs with 32 GB or more. Use the interactive Local AI Hardware & RAM Calculator to check exact disk, resident RAM, and KV-cache fit for your Mac or PC.
You do not need the largest model to record or transcribe a meeting. Choose a compact speech model and a smaller summarizer first, then move up only if the quality gain is worthwhile on your hardware.
Plan disk space on a Mac
- Compact setup: a roughly 0.7 GB Qwen3-ASR speech model plus a small 0.53–1.28 GB language model.
- Broader transcription: Whisper large-v3 turbo is roughly 1.6 GB; Qwen3-ASR 1.7B is roughly 2.5 GB.
- Quality-focused summaries: larger catalog choices add about 5–18 GB each.
- Meetings and day memory: audio, transcripts, and visual captures use additional space over time.
These numbers describe model downloads and may change when publishers update packaging. LokalBot lets you choose rather than downloading every model.
macOS permissions
Grant only the permissions for features you intend to use:
- Microphone records your voice.
- System audio / Screen Recording enables the selected meeting-app audio capture and visual screen context.
- Calendar helps detect and title scheduled meetings.
- Accessibility supports meeting detection, Cotyping, dictation insertion, and approved agent interaction.
On a fresh install, LokalBot asks through a notification before it records a detected meeting, and day tracking saves only app names and window titles. Visible text and encrypted screenshots are opt-in, gated by macOS permissions, and deleted after 14 days by default.
Windows (alpha)
- Operating system: Windows 10 or 11 on an x64 PC.
- Install: download the Windows installer from the latest release and run it. It installs LokalBot for your user into
%LOCALAPPDATA%\Programs\LokalBot, with a Start menu shortcut and no administrator rights. It isn't code-signed yet, so SmartScreen asks first: choose More info → Run anyway. The release also has a portable zip. - System audio: WASAPI process loopback records just the meeting app on build 20348 or later, which includes Windows 11. On older builds, or when the app can't be separated, WASAPI loopback records everything the computer plays.
- Disk: about 4 GB for the default transcription, summary, and search models, plus your recordings.
- GPU: optional. The built-in llama.cpp runs on a Vulkan GPU or on the CPU.
- Library:
%LOCALAPPDATA%\LokalBot.
Linux (alpha)
- System: x86_64 with glibc 2.39 or later, such as Ubuntu 24.04, Debian 13, or Fedora 40 or newer.
- Graphics: a Vulkan or OpenGL driver. Mesa is fine.
- Audio: PulseAudio or PipeWire. LokalBot records the meeting app's own streams where it can, and otherwise the output monitor, which captures everything the computer plays.
- Install: download the Linux archive from the latest release, extract it, and run
./install.sh. It installs LokalBot to~/.local/binand adds it to your applications menu. The same release also has a.deb(Debian, Ubuntu), an.rpm(Fedora, openSUSE) and an Arch package. Installing the.deb, or the.rpmon Fedora, also adds the LokalBot repository, sosudo apt upgradeorsudo dnf upgradebrings later versions; with the Arch package, LokalBot tells you about updates instead of installing them. - Disk: about 4 GB for the default models, plus your recordings. The built-in llama.cpp runs on a Vulkan GPU or on the CPU.
- Screen text in other scripts: needs Tesseract (
tesseract-ocron Debian and Ubuntu,tesseracton Fedora and Arch), which the.deband.rpmrecommend. Without it, screen text reads Latin letters with ocrs. - GNOME: the tray icon needs the AppIndicator extension, and day memory needs the Window Calls extension. Autocomplete in other apps needs LokalBot added as an input source, and in Firefox and Chrome it also needs Window Calls.
- Library:
~/.local/share/LokalBot.
X11 and Wayland
On Linux, a few features depend on whether your desktop session runs X11 or Wayland:
| Feature | X11 | Wayland |
|---|---|---|
| Day memory (app and window activity) | Yes | Hyprland, Sway, and KDE; GNOME with the Window Calls extension |
| Screen context (visible text, optional screenshots) | Yes | Sway and Hyprland with grim; other desktops through the screenshot portal |
| Dictation | Hold Alt+Space | Bind lokalbot dictate to a shortcut in your desktop's settings |
| Selection actions | Press Ctrl+Alt+Space | Bind lokalbot selection-actions to a shortcut in your desktop's settings |
| Autocomplete in other apps | Yes | As an input method on Sway, Hyprland, GNOME, and KDE Plasma, lightly tested so far (see the Linux guide); elsewhere only in the practice editor in Settings |
Models on Windows and Linux
The Windows and Linux app has its own defaults:
- Transcription: Qwen3-ASR 0.6B, with Parakeet v3 and v2, SenseVoice, GigaAM for Russian, and Whisper turbo as options.
- Speakers: pyannote segmentation with TitaNet embeddings.
- Autocomplete: Gemma 4 E2B, the same as on a Mac, or LFM2.5 1.2B with the Lightweight local preset.
- Summaries and search: Qwen3.5 4B and EmbeddingGemma 2, the same as on a Mac.
- Screen text: accessibility (UI Automation on Windows, AT-SPI on Linux) plus on-device OCR in your app and system languages: the recognizer built into Windows, or Tesseract on Linux, with ocrs as the fallback for Latin letters. Any download starts only when you ask.
- Spoken summaries: Kokoro 82M.
The benchmarks were measured on a Mac, so these engines have not been benchmarked yet.
Mac only for now
The Windows and Linux app has the same pages and reads the same library, but these features haven't been ported yet:
- Gmail in meeting briefs
- Apple Intelligence
- Granite Speech transcription
- Nemotron speaker separation
What works differently on Windows and Linux
- Remembered speakers: later meetings offer a remembered name for you to apply. Unlike on a Mac, nothing is named automatically, because the matching hasn't been calibrated on real calls with this voice model.
- Selection actions: the shortcut is Ctrl+Alt+Space, because ⌃⇧ W closes tabs and windows there. Apps without accessibility get results you can copy but not replace in place.
- Files: scanned PDF pages are read only when stored as JPEG or plain raster,
.docand RAW files aren't read, and video frames, HEIC, and AVIF need ffmpeg. Results open in their default app. - Phones and tablets: they pair as they do with a Mac. A firewall must allow TCP port 17890.
Phones and tablets
- iPhone and iPad: iOS or iPadOS 18 or later. Install the app from the App Store.
- Android: Android 10 or later. Install the app from Google Play, or the APK attached to each release. They're signed with different keys, so Android won't update one with the other: switching means uninstalling first, which removes the phone's copy and its pairing.
- A computer to pair with: LokalBot on a Mac, Windows, or Linux, with Settings → Phones & tablets turned on. The phone reaches it over your local network or Tailscale, with no LokalBot server in between.
Before an important meeting
- Install the latest release and finish the selected model downloads.
- On a Mac, confirm the required macOS permissions in System Settings.
- Make a short recording in the actual meeting application.
- Play back both the microphone and system-audio tracks.
- Process the recording once, offline if offline use matters.