LokalBot

LokalBot

Your personal context engine.

Notes from every call with no bot. One search for everything you saw and heard. Autocomplete and dictation in any app. Local by default. Your AI agents can search it too, from Claude Code, Codex, or any MCP client.

The real app on fictional data. Waits are cut and some typing is sped up.

Free and open source. No account.

Apple Silicon (M1 or later) with macOS 15+ · Windows 10/11 x64 (alpha) · Linux x86_64 (alpha) · on Mac, also brew install --cask stevyhacker/tap/lokalbot

⭐ If LokalBot is useful to you, star the repo. It's built by one developer, and stars are how other people find it.

English · 简体中文

Features · Agents & CLI · Windows & Linux · Compare · Privacy · Get started · Models · Website

Features

Remember what was agreed

Record microphone and meeting-app audio without a bot joining. Get local transcripts, recaps, decisions, and action items, with citations back to the discussion. You can also import audio and video files as meetings.

Find something you saw

One search across your meeting transcripts and, if you turn it on, what you saw on screen. Search by keyword or by meaning, then open the matching passage or screen moment instead of hunting through apps. The same window (⌃⇧ Space, or Ctrl+Shift+Space on Windows and Linux, once you turn on the shortcut in Settings → General) also runs app commands such as Record now or Start dictation from whatever app you're in.

Finish the sentence

As you type in other apps, gray text suggests how your sentence ends. Press Tab to take the next word, ` to take the rest, or keep typing.

• Knows what you're answering. It reads the message or page above the field, so a reply fits the thread.
• Remembers your meetings, if you let it. Turn on meeting and work memory, and it can bring in what was decided, like the ship date above.
• Writes in your language. On Mac, the same small model keeps up in German, French, Italian, Spanish, and more.
• Stays on your computer. A small model runs inside the app: Gemma 4 E2B on Mac, LFM2.5 1.2B on Windows and Linux. What you type and the context it reads never leave your computer.

Turn it on in Settings → Autocomplete, where you can try it on a sample conversation first and later see how much it wrote for you.

Pick up where you left off

Ask, then check the source

Ask your own work a question. Answers cite their sources, and say so when one can't be verified. Follow a citation back to the transcript or audio. Generated notes are a starting point, not a substitute for the discussion.

Talk instead of typing

Turn on Dictation in Settings → Dictation, then hold ⌥ Space (Alt+Space on Windows and Linux, or a shortcut you choose there), speak, and release to insert text at your cursor in any app. Transcription runs on your computer, and your recent dictations are kept so you can copy one again. In Compose mode, say what you need and LokalBot writes it at your cursor, from your meetings if you let it. On Wayland, bind lokalbot dictate to a shortcut in your desktop's settings.

Rewrite what you selected

Turn on selection actions in Settings → General, select text in any app, and press ⌃⇧ W. Fix the grammar, make it shorter, translate it, summarize it, or type your own instruction, then replace the selection or copy the result. Draft a reply with my memory also searches your meetings, so the reply can use what was actually agreed, and lists the meetings it used. It runs on your Think model, which stays on your computer unless you chose and approved a server. On Windows and Linux the shortcut is Ctrl+Alt+Space, because ⌃⇧ W closes tabs and windows there; on Wayland, bind lokalbot selection-actions to a shortcut in your desktop's settings.

It remembers how you work

Personal Context is an encrypted list of your facts, preferences, goals, and principles. Tell Ask "Remember I want the agenda the day before," add items in Settings, or let the overnight review pick up facts and preferences from your own words in meetings. Ask uses what it holds, so answers fit your work.

• Only your own words. It learns from your lines, never other speakers'. The model that reads them stays on your computer unless you approve a server.
• You see where it came from. Every learned item shows the line behind it. Preferences, and anything from what you typed, wait for you to keep or dismiss them. Facts and goals from meetings are saved directly.
• Yours to turn off. Learning is on by default, at up to three items a night. One switch in Settings stops it.
• Private from agents. Agents read only the items you choose to share, under a separate permission.

Draft the follow-up

Prepare local follow-up drafts. Also keep a note for every meeting in your Obsidian or Logseq folder, and get daily briefs and weekly work logs. You review and send them yourself.

The clips show the real Mac app on fictional data, with waits cut and some typing sped up. Search, Today, Timeline and People use a seeded library; the notes, Ask answers, dictation, autocomplete and Files results come from the on-device models. How they were made.

Also on Windows and Linux (alpha)

A native Windows and Linux app, written in Rust, has the same pages and reads the same library, so a library moves between computers. Encrypted items, such as Ask history, Personal Context, screenshots, and remembered voices, open only on the computer that saved them. Meetings, transcription, Ask, search, day memory, screen context, dictation, autocomplete, selection actions, spoken summaries, Files, overnight review, Personal Context, calendar, Agent Mode, and the CLI and MCP server all work. It's an alpha: the Windows build isn't code-signed yet, some features depend on your Linux desktop (Wayland support is partial), and Gmail in meeting briefs hasn't been ported. See what works where and screenshots.

For developers & agents

Claude Code, Codex, or any MCP client can search your meetings through the bundled lokalbot-cli or its stdio Model Context Protocol (MCP) server. Ask what was decided, who owns an action item, or where a topic came up. Agents can read your library but can't change it.

Connect Claude Code with one command, or install the Claude Code plugin:

# macOS
claude mcp add lokalbot -- /Applications/LokalBot.app/Contents/Helpers/lokalbot-cli mcp
# Windows and Linux, with lokalbot-cli on your PATH
claude mcp add lokalbot -- lokalbot-cli mcp

Access is off by default. On a Mac, turn on Allow external agents to read your LokalBot data in Settings → External Agents and tick what agents may read: Meetings, Screen memory, What I shared about myself, or Files. On Windows and Linux, each has its own switch under External agents in Settings → Privacy & Data: meeting library, files, screen memory, and What I shared about myself. Then install the CLI from Settings → Advanced → Agent CLI.

Ready to act on a meeting? Open any action item in Claude Code or Codex, and the prompt is already typed in with the task and where it came from, for you to review and send.

Or use the CLI from any shell or agent:

lokalbot-cli search "database decision"
lokalbot-cli get latest --include metadata,summary
lokalbot-cli search-files "lease renewal" --kind pdf   # with Files access on
lokalbot-cli mcp

Without a PATH installation, use /Applications/LokalBot.app/Contents/Helpers/lokalbot-cli on Mac. On Windows and Linux, Settings → Advanced → Agent CLI installs it too; see Coding agents for where it goes and how to add it to your PATH. The commands and MCP tools are the same on every platform.

Screen-memory MCP tools require separate permission scoped to today, the last seven days, or all retained history. They return text and metadata, not decrypted screenshots. search_files searches the folders you chose under Files by their words and returns file paths, excerpts, and the matched passages; it needs Files access and never changes your files. External clients may send results to their own model providers.

The app also has an optional Agent Mode for working with reviewed context. File and shell actions follow its approval settings; it is separate from the read-only CLI/MCP interface.

CLI examples · Claude Code plugin · MCP architecture

How it compares

Coming from…	What's different with LokalBot
Granola	The same bot-free capture, but transcription and summaries run on your computer by default. No account and no subscription.
Apple Notes	Notes records your microphone, so with headphones on it misses the other side of a Zoom or Teams call. LokalBot records both sides and keeps every meeting searchable, with notes that cite the discussion.
Otter · Fireflies	Nothing joins your call, audio stays on your computer, and there are no minute caps.
Hyprnote (anarlog)	Local AI ships built in, with no separate model runner or API keys to set up. Dictation, a day timeline, and autocomplete come in the same app.
Meetily	Both are open source and run locally. Speaker separation, meeting detection, and MCP access are free here rather than Pro features, and dictation, a day timeline, and autocomplete come in the same app.
Screenpipe	Starts with app and window activity. Screen text and screenshots are opt-in and deleted after 14 days by default. GPLv3 rather than source-available.
Rewind	Rewind's Mac app is discontinued. LokalBot's timeline is open source and actively developed.
MacWhisper · Superwhisper	Dictation comes alongside meeting notes, recall, and autocomplete, for free. For transcribing files you already have, MacWhisper is the better tool.
Wispr Flow	Dictation and bot-free meeting notes run on your computer, with no account, word cap, or connection required. Wispr Flow transcribes both in its cloud.

Competitor details were last verified on September 24, 2026, and October 1, 2026 for Apple Notes, Meetily, and Wispr Flow; each link has the full comparison and sources. Trademarks belong to their owners.

Privacy

Your work is personal. Keeping it local should be the starting point.

Recording and built-in AI processing happen on your computer. LokalBot has no account system, analytics service, or telemetry backend. Your library lives in local files and SQLite under your user account.

flowchart LR
  subgraph mac [Your computer]
    direction LR
    audio["Mic + meeting-app audio"] --> asr["On-device<br/>transcription"]
    asr --> llm["Local summaries<br/>built-in llama.cpp"]
    screen["Opt-in screen text"] --> db[("Local library<br/>full-text + vector search")]
    asr --> db
    llm --> db
    db --> use["Search · Ask · Today · Timeline"]
    db -.->|read-only, off by default| cli["lokalbot-cli · MCP"]
  end

• You choose what to retain. Activity-only tracking is the default. Visible screen text and encrypted screenshots are opt-in. Pause capture or exclude apps and domains at any time. Detection has limits; exclude anything you never want retained.
• You control retention. Captured screen text and screenshots expire after 14 days by default. Adjust the window or explicitly save a moment to keep it until you unsave or delete it.
• Network access has boundaries. Models, updates, and optional agent runtimes need downloads. Automatic update checks can be disabled. A remote model server receives context only after you approve its origin; optional Agent Mode and external CLI/MCP clients have separate permissions and data handling.
• Your phone gets a copy only if you pair it. Settings → Phones & tablets is off by default. When you turn it on, the Mac sends the phones and tablets you pair an end-to-end encrypted copy of the data you choose, over your local network or Tailscale, with no LokalBot server in between. Screenshots, screen text, attendee emails, and voice profiles are never shared. A paired phone can also record a meeting and send it the same way; your Mac processes it, and the phone then deletes its copy.

To check the local processing path, download the models, select the built-in backend, disable automatic update checks, and observe network traffic while recording, transcribing, and summarizing. Approved agent commands and external clients are separate paths.

Read the full privacy policy · Report a security issue privately

Get started

LokalBot runs on Apple Silicon Macs with macOS 15 or later. Windows and Linux are in alpha. You'll also need space for your selected models: Settings → Models shows download sizes and readiness, and storage and memory needs depend on the models you choose.

	macOS	Windows (alpha)	Linux (alpha)
Runs on	Apple Silicon (M1 or later), macOS 15+	Windows 10 or 11, x64	x86_64 with glibc 2.39 or later (Ubuntu 24.04, Debian 13, Fedora 40 or newer), PulseAudio or PipeWire, and a Vulkan or OpenGL driver
Install	LokalBot-<version>-macos-arm64.dmg: drag LokalBot to Applications. Or brew install --cask stevyhacker/tap/lokalbot.	LokalBot-<version>-windows-x86_64-setup.exe: run it to install LokalBot for your user, with a Start menu shortcut and no administrator rights. It isn't code-signed yet, so SmartScreen asks first: More info → Run anyway. The same release has a portable zip.	LokalBot-<version>-linux-x86_64.tar.gz: extract it and run ./install.sh, which adds LokalBot to your applications menu. Or install the .deb (Debian, Ubuntu), .rpm (Fedora, openSUSE) or Arch package from the same release. The .deb, and the .rpm on Fedora, add the LokalBot repository, so sudo apt upgrade or sudo dnf upgrade brings later versions; with the Arch package, LokalBot tells you about updates instead of installing them.
Your library	~/Library/Application Support/me.dotenv.LokalBot	%LOCALAPPDATA%\LokalBot	~/.local/share/LokalBot

A library recorded on one platform opens on the others. Not every Mac feature is on Windows and Linux yet; see what works where.

1. Install the app for your platform and open it.
2. Choose what to capture. Review recording settings: automatic detection, ask-first, or manual recording. Day tracking starts with activity only; saving screen text or screenshots is a separate opt-in. Grant only the permissions for the features you enable.
3. Let the models download. Built-in transcription, search, summaries, and writing can then work offline. No model server to set up and no API key to bring.

Try this first: make a short test recording, let it transcribe, then search for a phrase you said and open the matching passage. Add screen memory or writing tools when you're ready.

Let meeting participants know before recording and get their consent.

Release notes · Setup help · Homebrew options · Windows and Linux guide

FAQ

No. LokalBot records your microphone and the meeting app's audio on your computer. On Windows and Linux it records the meeting app alone where it can, and otherwise everything the computer plays. Audio sources are separate from individual speaker identification. When system-audio capture is unavailable, the app warns that recording is microphone-only. Always record with participants' consent.

No. New installs use activity-only tracking, which you can turn off. Saving visible text or screenshots requires a separate choice and the relevant macOS permissions (on Linux, your desktop's accessibility and screenshot support). Excluded apps and domains and secure fields are skipped; private browser windows are captured unless you exclude the browser or site. Detected credentials are redacted and associated pixels are dropped, but detection is not perfect. See PRIVACY.md.

Yes. Turn on Meeting note export to keep a Markdown note for every meeting in a folder you choose, such as an Obsidian vault or a Logseq graph: the summary, action items, the people and projects involved, and the transcript. You can also export a single meeting, write a daily memory note, or let local routines save drafts. These are ordinary, unencrypted files; any service you sync them with has its own privacy settings. See Meeting transcripts in Obsidian.

Yes, as an alpha. The Windows and Linux app is a native Rust app with the same pages, the same models for summaries and search, and the same library format, so a library moves between computers. Encrypted items, such as Ask history, Personal Context, screenshots, and remembered voices, open only on the computer that saved them. Gmail in meeting briefs is Mac only for now, and on Wayland some features depend on your desktop. See what works where.

Local AI, your choice

Start with the built-in models, then change them in Settings → Models. LokalBot supports its bundled llama.cpp runtime, Ollama, OpenAI-compatible servers, and Apple Intelligence on supported Macs running macOS 26 or later. Think can also run on your own ChatGPT plan (Sign in with ChatGPT) or Claude plan (through your signed-in Claude Code), without an API key, and Agent Mode can run on the ChatGPT plan. Remote servers and plans require approval before receiving context.

Different jobs use different models. These are the Mac defaults on a fresh install, about 9.6 GB of model files if you use every role. They are not a RAM requirement:

Role	Default model	Format	Approx. model files
Transcription	Qwen3-ASR 1.7B	MLX 8-bit	2.47 GB
Summaries and chat	Qwen3.5 4B	Q4_K_M	2.74 GB
Autocomplete	Gemma 4 E2B Base	Q6_K	3.85 GB
Semantic search	EmbeddingGemma 2	Q8_0	0.31 GB
Speaker diarization	Nemotron 3 (preview) via FluidAudio	Core ML	~0.20 GB

Qwen3-ASR covers 52 languages and dialects. Transcription alternatives include Parakeet, Granite Speech, Whisper large-v3 turbo, SenseVoice for Chinese, Japanese, and Korean, and GigaAM for Russian; Pyannote Community-1 remains available for diarization. The Lightweight preset swaps in smaller models, including LFM2.5 1.2B Instruct (0.73 GB) for autocomplete. Model choices can vary by release; Settings shows the active selections and available presets. Speed and memory use depend on your computer, model, context length, and other workloads.

On Windows and Linux, transcription uses Qwen3-ASR 0.6B through sherpa-onnx (Parakeet v3 and v2, SenseVoice, GigaAM Russian, and Whisper turbo are optional), speakers are separated with pyannote segmentation and TitaNet, and autocomplete uses LFM2.5 1.2B by default. Summaries (Qwen3.5 4B) and search (EmbeddingGemma 2) use the same models as on Mac. The default download is about 4 GB, and llama.cpp runs on a Vulkan GPU or the CPU.

Project benchmarks · Model catalog · Search implementation · Model and runtime details

Build from source

For the Mac app, use an Apple Silicon Mac with full Xcode installed; the build workflow pins the CI toolchain. With Homebrew available:

brew install xcodegen cmake
# Rust (https://rustup.rs) builds the phone and tablet sync library; the repo
# pins the toolchain, which rustup installs on first use.
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y --default-toolchain none

git clone https://github.com/stevyhacker/lokalbot.git
cd lokalbot

# XcodeGen needs these paths before the first build fetches the runtimes.
mkdir -p Vendor/llama-cpp Vendor/sherpa-onnx
# The phone and tablet sync library (Rust), as a local Swift package.
Scripts/build-sync-xcframework.sh
xcodegen generate
open LokalBot.xcodeproj

Select LokalBot Dev, set your signing team, and run. The Dev app has its own identity, macOS permission grants, library, model storage, and Keychain namespace, so it can live alongside the installed release without applying development retention settings to release data. The first build prepares pinned native runtimes; model downloads happen separately.

For the Windows and Linux app, see Build from source in its README (Rust).

Build and test commands · Testing guide · Screenshot guide

LokalBot/
├── Views/       # Native SwiftUI screens, settings, and onboarding
├── Services/    # Recording, search, storage, and workday memory
├── Engines/     # Transcription, model catalog, and local inference
├── Cotyping/    # System-wide autocomplete
├── Dictation/   # Voice typing
└── Agent/       # Optional Agent Mode and approval flow
CLI/             # Command-line interface and MCP entry points
LinuxWindows/    # Windows and Linux app (Rust + GPUI)
LokalBotTests/   # Unit tests
Scripts/         # Build, verification, capture, and release tooling
project.yml      # XcodeGen source of truth

Contributing

Bug reports, documentation improvements, and code contributions are welcome. Report an issue with steps to reproduce, expected and actual behavior, the app version, your OS and its version (on Linux, also the distribution, desktop, and X11 or Wayland), your chip or CPU and GPU, RAM, and selected models. Keep private recordings and transcripts out of reports.

For larger changes, open an issue to discuss the approach first. Follow the working agreements and development guide, keep pull requests focused, and record your checks in the PR template. Include screenshots for UI changes; run UI tests on hosted CI or a remote runner.

License & acknowledgements

LokalBot is free software under GPLv3. Code in the Autocomplete engine that was ported from Cotabby stays under AGPLv3. Other third-party components and models retain their own licenses; see third-party notices.

Built with llama.cpp, MLX Swift, speech-swift, sherpa-onnx, WhisperKit, FluidAudio, Sparkle, XcodeGen, GPUI, GPUI Kit, and ONNX Runtime, with models including Qwen3-ASR, IBM Granite Speech, Parakeet, and LFM2.5. The Autocomplete engine started from Cotabby's open-source code.