# screenpipe > AI agent memory. Local-first 24/7 capture of screen and audio, plus an open API any agent can query. > Full reference: https://screenpipe.com/llms-full.txt Screenpipe is a local-first, private AI agent memory that captures what you've seen, heard, and done, giving AI agents the context and information they never had, so they never start from scratch. screenpipe runs in the background on macOS, Windows, and Linux. It records every screen and every sound on the device, extracts text via OCR and audio via on-device speech-to-text, and stores everything in a local SQLite database. From there it exposes a local REST API + MCP server that AI agents (Claude Code, Cursor, ChatGPT, custom pipes) can query for full context of what the user is doing, without the user having to re-explain their work history. The thesis: every AI agent needs to know what the user is doing, act on that context, and trigger without prompts. Vision is the universal interface; the screen is the highest-fidelity signal of human intent. screenpipe is the context infrastructure between the two. Open source, source-available. Local-first by design. Data never leaves the device unless the user explicitly opts into cloud sync or a hosted model. 19,000+ GitHub stars. ## What screenpipe does - **Record**: continuous screen capture (all monitors), system audio, and microphone, 24/7, in the background - **Rewind**: visual timeline of everything that happened on the device, queryable by app, window, URL, or natural language - **Ask**: natural-language search across OCR text and transcripts, with screenshots and audio clips returned alongside text - **Pipes**: extensible plugin system: TypeScript/Bash scripts that react to screen activity (meeting summaries, daily journals, CRM sync, custom workflows) - **API + MCP**: local HTTP API and Model Context Protocol server, so any AI agent can query the user's context ## How it differs from alternatives - **vs Rewind / Limitless**: open source, cross-platform (Rewind/Limitless are Mac-only), all data stays on-device by default - **vs Microsoft Recall**: works today on macOS, Windows, and Linux; not gated to Copilot+ PCs; open source and auditable - **vs Granola / Otter / Fireflies / tl;dv**: not just meetings, captures the entire desktop continuously, including browsing, docs, messaging, and code - **vs ChatGPT / Claude memory**: server-side memory features only remember chat content; screenpipe captures everything you actually do - **vs proprietary "memory" SaaS**: data stays on-device; no per-seat lock-in; build custom workflows via the API and pipe SDK Compare pages: https://screenpipe.com/compare ## Who it's for - **Developers and AI engineers** building agents that need context of what the user is doing, via Claude Code / Cursor / Cline integrations, custom MCP servers, autonomous workflows - **Knowledge workers** who want their AI assistants to know what they were doing five minutes ago without describing it - **Enterprises** with compliance / data-residency / sovereignty requirements that rule out cloud memory tools - **Researchers and indie hackers** who want a programmable layer over their own digital life ## Technical surface - **Platforms**: macOS (Apple Silicon + Intel), Windows 10/11, Linux - **Capture**: native OCR (Apple Vision on macOS, Windows OCR API, Tesseract on Linux); on-device transcription (Parakeet-MLX on Apple Silicon, Whisper.cpp elsewhere; optional Deepgram cloud) - **Storage**: local SQLite + WAL; configurable retention; ~5-10 GB / month typical - **API**: REST on `localhost:3030` plus an MCP server consumable by Claude Desktop, Cursor, Continue, etc. - **Pipes SDK**: write plugins as `pipe.md` files with embedded `bash` / `bun` blocks; runs sandboxed in the user's environment - **Performance target**: <20% CPU, <3 GB RAM on release builds, designed to run forever in the background ## Links - [Homepage](https://screenpipe.com/): product overview - [Onboarding / Download](https://screenpipe.com/onboarding): get the app - [Compare](https://screenpipe.com/compare): vs Rewind, Limitless, Microsoft Recall, Granola, Otter, tl;dv, etc. - [Claude Code Memory](https://screenpipe.com/claude-memory): how Screenpipe gives Claude Code persistent memory via MCP - [Pipes](https://screenpipe.com/pipes): plugin marketplace - [SDK](https://screenpipe.com/sdk): building on top of screenpipe - [Enterprise](https://screenpipe.com/enterprise): on-prem / SSO / data-residency - [Privacy Filter](https://screenpipe.com/privacy-filter): PII redaction layer (Tinfoil) - [Security](https://screenpipe.com/security): threat model, encryption, audit posture - [Changelog](https://screenpipe.com/changelog): releases - [Research](https://screenpipe.com/research): technical posts - [Blog](https://screenpipe.com/blog): announcements and tutorials - [Privacy Policy](https://screenpipe.com/privacy) - [Terms](https://screenpipe.com/terms) ## Source code and community - GitHub: https://github.com/screenpipe/screenpipe - Documentation: https://docs.screenpi.pe - Discord: https://discord.gg/screenpipe - Twitter / X: https://twitter.com/screen_pipe - Contact: louis@screenpi.pe ## Licensing MIT for the core engine, CLI, and most of the desktop app. Proprietary cloud services (sync, hosted models) are gated behind a Pro tier; everything that runs locally is free and source-available.