Purpose: screenpipe continuously captures screen and audio locally, forming a searchable computer work history, used to recall what was seen and heard, map workflows, provide context for AI agents, and generate automatable tasks.
Main capabilities: event-driven screenshots, taking screenshots only on meaningful changes such as app switches, clicks, typing pauses, and scrolling, with accompanying accessibility tree text and a fallback to OCR when unavailable; audio transcription supports system audio and microphone, local Whisper (Large-V3-Turbo) or cloud Deepgram, including speaker recognition and diarization; AI natural language search, filterable by app name, window title, browser URL, and date, backed by SQLite FTS5; visual timeline, allowing screenshots and audio to be replayed like a DVR; Pipes plugin system, defining scheduled AI agents in markdown, with built-in meeting summaries, daily reviews, standup updates, time breakdowns, AI prompt logs, video export, and more; MCP server, allowing clients such as Claude Desktop, Cursor, and VS Code to query screen history; local localhost REST API (default port 3030) and JavaScript/TypeScript SDK.
Usage: download the macOS, Windows, or Linux desktop app from the official website, or use the CLI. To install MCP, run: claude mcp add screenpipe -- npx -y screenpipe-mcp@latest. Developers can create a pipe by writing markdown in ~/.screenpipe/pipes/, and can restrict accessible apps, windows, content types, time ranges, and endpoints via YAML frontmatter.
Suitable scenarios: knowledge workers recalling content, developers providing context to coding assistants, researchers retrieving screen information, remote workers automatically recording meetings, teams and enterprises deploying centrally.
Limitations: local capture and search can be used offline after setup, but cloud AI, sync, and connection services require a network; resource usage reference is about 5-20% CPU and about 0.5-3 GB memory (varies by hardware and settings), storage varies with activity level, the official recommendation is 8 GB memory and measure it yourself; Linux capture behavior depends on distribution, desktop environment, and permissions; source code is available under a license permitting personal non-commercial use, commercial use requires a commercial license. For pricing and specificpackage contents, please refer to the current page on the official website.
