---

_{A Monday, as it was kept. Every tile is a screen the engine held on to, and the panel is the text it read off the one selected.}

All work and no play makes Jack a dull boy.

You work all day and your agents still start from scratch.

litepipe keeps a local memory of everything you do, hear, and see on your computer: meetings, videos, clicks, every app. It all lands in one folder on your Mac.

Screenpipe built a solid capture engine and published it under MIT. litepipe is a fork that keeps that core evolving in the open, without the features that grew around it: no pipes, no accounts, no telemetry, no cloud sync. The work goes into simplicity, less weight, and safer guardrails on what gets captured. Bugs in the shared code go back upstream, as in issue 5531 and pull request 5532.

Features

• Meetings become transcripts. litepipe spots the call window and asks. One click, and the transcript is on your disk when the call ends.
• The microphone opens only for meetings you accept. The recording is deleted once it's text.
• Everything on screen is captured. Text comes from the accessibility tree, with OCR for what the tree can't see.
• Nothing leaves the Mac. Whisper and pyannote run on device, and the only socket is the app talking to its own engine.
• One folder holds it all. SQLite with full text search, so your agent reads it directly.
• Any agent can read it. A bridge ships inside the app, so Claude Code, Claude Desktop, Cursor or anything else that speaks MCP connects with one line and no install. Read only, no port, no key. See docs/MCP.md.

Install

App

Download litepipe.dmg, drag it to Applications, launch. Signed and notarized by Apple.

Source

To audit or change the code. Needs Xcode. The engine ships prebuilt; rebuilding it needs Rust and CMake, see crates/.

git clone https://github.com/Rafaelpta/litepipe
cd litepipe/apps/litepipe-mac
./build-app.sh release

CLI

The engine alone, no app. Captures into ~/.litepipe, local API on 127.0.0.1:3030, control C stops it. No meeting banner, no microphone gate, and the permissions go to your terminal.

curl -fsSL https://raw.githubusercontent.com/Rafaelpta/litepipe/main/headless.sh | bash

How it works

The engine reads what you're working on through the accessibility tree, the layer assistive technology uses: like HTML, for every app. OCR fills the gaps the tree can't see: video, games, remote desktops. Voice is transcribed locally with Whisper and grouped by speaker. Everything lands in SQLite with full text search, so you can query it, search it, or point your AI agent at the folder and ask what you agreed to, planned, or missed.

Screen and voice take separate paths, both end as searchable text in the same database, and everything is redacted before it settles.

flowchart LR
    SCR["Screen<br/>all day"] --> STT["Screen to text<br/>accessibility tree + OCR"]
    MIC["Voice<br/>meetings you accept"] --> VTT["Voice to text<br/>Whisper, after the call"]
    STT --> PII["PII redaction<br/>seconds later, overwritten in place"]
    VTT --> PII
    PII --> DB[("~/.litepipe/db.sqlite<br/>full text search")]
    VTT -. "audio deleted once transcribed" .-> X["no voice file remains"]

Privacy controls

Capture isn't all or nothing. What ships today:

• Private windows are never captured. Safari, Chrome, Edge, Brave, Arc, and Firefox.
• Password managers are never captured. 1Password, Bitwarden, LastPass, Dashlane, KeePassXC, and Keychain Access.
• Any app or site can be excluded. Settings, Privacy takes two lists, one of websites and one of apps. An excluded window never reaches the capture buffer, so no frame and no text of it exists.
• Capture stops when you want. One shortcut pauses everything, hours can be set to a schedule, and DRM video pauses it on its own.
• Secrets are redacted. On by default: keys, cards, and passwords become labels in text and black boxes in screenshots, and the original is overwritten. Under the hood: 46 deterministic patterns run first, an ONNX model on the Apple Neural Engine catches what patterns can't, and a second model finds secrets in the pixels and paints them solid black rather than blurred, since a blur can be undone.
• What was captured can be deleted. Settings, Data shows how much litepipe holds and deletes the last hour, the last day, a period you pick, or everything, files included. The local API does the same for scripts.
• The whole memory can be locked. The vault encrypts database and media with a key derived from your password, held only on your machine.

Architecture

Everything runs in two processes on your machine, and what they may write is checked twice: before capture, and before it settles on disk. Every claim below can be checked from a shell on your own machine.

flowchart LR
    IN["Screen, screen text, system audio<br/>Microphone only in meetings you accept"] --> G1
    G1{"Checked before capture<br/>private windows, password managers,<br/>your app and site lists, pause, schedule"}
    G1 -->|"blocked"| NONE["Nothing is written<br/>no frame, no text"]
    G1 -->|"allowed"| CAP["Captured and transcribed<br/>on this machine"]
    CAP --> G2["Checked after capture<br/>secrets redacted in text and pixels<br/>voice audio deleted once transcribed"]
    G2 --> DISK[("~/.litepipe<br/>db.sqlite, media, logs")]
    DISK --> APP["litepipe.app"]
    DISK --> AGT["your AI agent"]

Processes and permissions

The app spawns the engine with posix_spawn and disclaims responsibility transfer, so it stays the TCC responsible process: the permission prompts and grants belong to litepipe.app, and one set covers both processes. Kill the app and the engine goes with it.

macOS permissions used: Screen Recording (capture), Accessibility (screen text and UI events), Microphone (only while a meeting you accepted is running).

What's written to disk

Everything lives in one folder, ~/.litepipe. Browse it with open ~/.litepipe.

Nothing is encrypted at rest by default: FileVault covers the disk, and the vault (/vault/*) locks the database and media behind a password when you turn it on. Settings, Data shows the folder's size and deletes by period, from the last hour to all of it. Removing ~/.litepipe by hand works too; the app rebuilds an empty one on the next launch.

Network

One listening socket: the engine on 127.0.0.1:3030, writes authenticated with a key generated at install time. The app is the only client; it calls meetings/start, meetings/stop, meetings/status, audio/start, audio/stop, and health, and reads history straight from SQLite. Confirm with lsof -nP -iTCP -sTCP:LISTEN | grep screenpipe or a network monitor.

Models are downloaded once on first run and then cached: Whisper and a voice activity model for transcription, and, while secret redaction is on, an image redactor (52 MB, ~/.screenpipe/models/rfdetr_v12.onnx) and a text redactor (159 MB, ~/.screenpipe/models/v45_phase5_pruned/).

Source map

Specs

FAQ

Does any audio or screen data leave my machine? No. The engine downloads its models on first run; after that the only network socket is the app talking to its own engine on 127.0.0.1. No telemetry, no crash reporting, no updater.

What happens to the recording of my voice? It's deleted within the hour and the transcript stays. A voice recording is the most sensitive file on the disk: minutes of it can clone a voice, and an hour of meetings is about 25 MB of audio against 50 KB of text.

Which meeting apps are detected? Zoom, Google Meet, and Microsoft Teams, native and web, plus FaceTime. Detection reads window titles and browser URLs from captured frames.

How do I keep something out of the memory? Four ways. Private browser windows are skipped on their own. Any app or site goes on the ignore list. The shortcut pauses everything. And anything already captured can be deleted by period in Settings, Data.

Is the data encrypted? FileVault covers the disk, on by default in macOS. The vault adds a lock on the database and media, behind a key derived from your password. Transparent encryption during capture is on the roadmap.

How much disk does it use? System audio is kept 7 days and screen frames 30 days, both cleaned up on their own. Text and the index stay, and they're small. Settings, Data shows the current size.

Can my AI agent read the data? Yes, that's the point. Plain SQLite and media files in ~/.litepipe. Point an agent at the folder and query it.

How do I delete everything? Settings, Data, delete all context. Or quit the app and remove ~/.litepipe.

Status

Early beta. Native macOS on Apple Silicon. Open an issue when something breaks.

Contributing

Bug reports and fixes are welcome. The direction is narrow on purpose: keep it local, keep it stripped to the basics, and make the capture engine better.

License

litepipe is MIT. It adapts open source code from Screenpipe and Clicky, both MIT, and bundles FFmpeg. The notices are in LICENSE.md, NOTICE, and THIRD_PARTY_NOTICES.md.