What it does screenpipe is a local-first screen and audio recorder that runs 24/7, building a searchable SQLite database of your digital life. It captures screenshots paired with accessibility-tree text (OCR as fallback), transcribes audio via local Whisper, and exposes it all through a REST API and MCP server so Claude, Cursor, or your own scripts can ask “what did I do in the last hour?”

The interesting bit The capture is event-driven, not brute-force: it listens for app switches, clicks, and typing pauses, then snaps only when something actually changes. That keeps CPU at 5–10% and storage around ~20 GB/month instead of drowning you in identical frames. The “pipes” system is clever too—scheduled agents defined in markdown files with YAML frontmatter permissions, so you can gate what AI sees by app, time of day, or data type without trusting a prompt.

Key highlights

• Event-driven screen capture + accessibility tree extraction, OCR fallback
• Local Whisper transcription with speaker diarization; per-app audio exclusions on macOS
• Natural language search across screen text, OCR, and transcripts with semantic embeddings
• MCP server for zero-config integration with Claude Desktop, Cursor, VS Code extensions
• Deterministic permission model for pipes: OS-level filtering by app, window, content type, time range, and raw SQL/frames gating
• Apple Intelligence support on compatible Macs for on-device summarization
• Cross-platform: native installers for macOS (Intel/Apple Silicon) and Windows; Linux builds from source

Caveats

• Linux requires building from source—no native installer yet
• ~20 GB/month storage adds up; budget your disk accordingly
• Some features (CoreAudio exclusions, Apple Intelligence) are macOS-only

Verdict Worth a look if you want Rewind.ai or Microsoft Recall without the cloud, and especially if you’re already using AI coding assistants that could use real context. Skip it if the idea of a full-time recorder on your machine—local or not—makes you flinch.