What works on Windows
LokalBot for Windows is a native app written in Rust with GPUI, the UI framework behind the Zed editor. It has the same pages as the Mac app and reads the same library format, so a library recorded on a Mac opens on Windows. It's an alpha.
- Meetings without a bot. LokalBot records your microphone and the meeting app's own audio through WASAPI process loopback on Windows build 20348 or later, which includes Windows 11, so nothing joins the call. On older builds, or when it can't separate the app, the second track is everything your PC plays. When it detects a meeting, it asks with a notification.
- On-device transcription. Qwen3-ASR 0.6B by default, with Parakeet v3 and v2, SenseVoice, GigaAM for Russian, and Whisper turbo as options, and speakers separated with pyannote segmentation and TitaNet.
- Notes and Ask. Summaries, decisions, action items, and answers with citations from Qwen3.5 4B in the built-in llama.cpp, Ollama, an OpenAI-compatible server you approve, or your ChatGPT or Claude plan.
- Your day. App and window activity, and, if you turn it on, screen text and encrypted screenshots, read through UI Automation and the text recognizer built into Windows, in your app and system languages where Windows has that language pack.
- Writing. Hold Alt+Space to dictate anywhere. Optional autocomplete suggests the next words in apps whose text fields support UI Automation, with Gemma 4 E2B on your PC, or the smaller LFM2.5 1.2B with the Lightweight local preset.
- Calendar. Private iCal or CalDAV links from Outlook, Google, or another calendar name your recordings and list today's meetings.
- Agents. A read-only CLI and MCP server for Claude Code, Codex, and other agents, and an optional Agent Mode that works on your files with your approval, in Git Bash or PowerShell.
- Selection actions. Turn them on in Settings → General, select text in any app, and press Ctrl+Alt+Space to fix the grammar, shorten it, translate it, or draft a reply from what your meetings say. Apps without accessibility get results you can copy but not replace in place.
- Spoken summaries. Hear summaries and answers read aloud, or export a WAV, with Kokoro 82M on your PC.
- Overnight review and Personal Context. LokalBot reviews your day overnight, waiting for mains power, and keeps the facts, preferences, goals, and principles you give it. Agents read only the items you choose to share.
- Files. Search folders you choose by words and meaning, off until you turn it on. 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. - Exports and coding sessions. Keep a Markdown note for every meeting and a daily note in your Obsidian or Logseq folder. Coding-agent sessions join day memory only after you choose that in setup or Settings → Day Memory.
- Phones and tablets. LokalBot can share your meetings and day with your own devices, off until you turn it on in Settings → Phones & tablets. Pair an iPhone or iPad with the app on the App Store, or an Android phone or tablet with the app on Google Play. A firewall must allow TCP port 17890.
Install without admin rights
Run LokalBot-<version>-windows-x86_64-setup.exe from the latest release. It installs for your user into %LOCALAPPDATA%\Programs\LokalBot, without administrator rights, and adds a Start menu shortcut and an Installed apps entry. Uninstalling keeps your library. A portable zip is in the same release.
The app isn't code-signed yet, so SmartScreen asks before the first launch: choose More info → Run anyway. Updates are signed: LokalBot checks for a new version once a day, verifies the download's signature, and installs it only when you choose Install and Relaunch. A winget package is in review.
Models and hardware
LokalBot needs Windows 10 or 11 on x64. On first launch it downloads the models you choose: about 4 GB for the default transcription, notes, and search models. llama.cpp runs on any Vulkan GPU, including AMD Radeon and Intel graphics, or on the CPU. On an NVIDIA GPU you can choose CUDA (experimental) in Settings → Models → llama.cpp; in testing on an NVIDIA L4, CUDA read a long prompt into Qwen3.5 4B 32% faster than Vulkan. See the system requirements.
What's not there yet
- Code signing, so SmartScreen will warn until it's in place.
- Gmail in meeting briefs, Apple Intelligence, Granite Speech transcription, and Nemotron speaker separation are Mac only for now.
- Remembered speakers only suggest. Naming a speaker teaches an encrypted voice profile, and later meetings offer that 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.
- Bot-free is not consent-free. Tell the people in your meeting that you're recording, and follow the rules that apply to them.