Screenpipe: 24/7 local screen memory your AI agents can query
The YC launch turns your workday into a searchable log, piped into Claude Code, Cursor and Cline over MCP — offline, local, source-available not free.
Copy markdown
Your screen becomes a searchable database
Screenpipe grabs a screenshot on OS events — app switch, click, typing pause — pairs it with the accessibility tree (OCR only as fallback), and indexes it into a local SQLite FTS5 store. It runs at ~5-10% CPU and about 20GB/month, fully offline.
Pipe it into the agents you already run
A REST API on localhost:3030 plus an MCP server let Claude Code, Cursor, Cline, Continue and Claude Desktop read your recent activity as context. There's a JS/TS SDK for wiring your own tools.
Local-first, PII redacted on-device
Frames, audio and transcripts stay on your machine; transcription runs via local Whisper/Parakeet and a built-in PII model redacts on Apple MLX or Windows DirectML at under 1% CPU. Optional zero-knowledge sync across devices, no account required.
Pipes: turn how you work into agents
'Pipes' are markdown + YAML-frontmatter agents with per-source permissions (app, window, time range). Ship a standup summarizer, a day recap, or an automation like 'LinkedIn edit to CRM update' this weekend.
The catch: source-available, not open source
The code is on GitHub but under a Screenpipe Commercial License — free for personal, nonprofit, education and research; paid for commercial use at $25-$150/seat per month. Pre-license builds stay MIT if you need permissive terms.