screenpipe

Purpose

Screenpipe is a source-available desktop app that continuously captures your screen and audio, creating a searchable, AI-powered memory of everything you do on your computer. It positions itself as a source-available alternative to tools such as Rewind.ai (now Limitless), Microsoft Recall, Granola, and Otter.ai, with an emphasis on local-first and auditability.

Main Capabilities

• Event-driven screen capture: Instead of recording every second, it listens for meaningful events such as app switches, clicks, typing pauses, and scrolling, and only takes a screenshot when the content actually changes. Each capture pairs the screenshot with the accessibility tree (the structured text the operating system knows about: buttons, labels, text fields); it falls back to OCR when accessibility data is unavailable (such as remote desktops or games).
• Audio transcription: Captures system audio (what you hear) and microphone input (what you say). Uses locally running Whisper (Large-V3-Turbo) for real-time speech-to-text, or Deepgram for cloud transcription. Supports speaker recognition and separation, and works with any audio source such as Zoom, Google Meet, and Teams.
• AI search: Natural language search over accessibility-first screen text, OCR fallback text, and audio transcriptions. Can filter by app name, window title, browser URL, and date range. Under the hood it uses SQLite FTS5 full-text keyword search, and results return both screenshots and audio segments.
• Visual timeline: Scroll through your day like a DVR, click any moment to view the full screenshot and extracted text, and replay the audio from any time period.
• Plugin system (Pipes): Pipes are scheduled AI agents defined in Markdown files. Each pipe is a pipe.md with a prompt and a schedule; Screenpipe runs an AI coding agent (such as pi or claude-code) to query screen data, call APIs, write files, and perform actions. Built-in pipes include: meeting-summary (summarize a meeting that just ended and write notes back), day-recap (today's accomplishments, key moments, and unfinished work), standup-update (what was done, next steps, blockers), time-breakdown (statistics on where time went by app, project, and category),ai-prompt-journal (captures every prompt you send to AI tools and saves them to Obsidian or local Markdown), video-export (generates videos of recent screen activity). Developers can create a pipe by writing a Markdown file in ~/.screenpipe/pipes/.
• MCP server: Screenpipe runs as an MCP server, allowing AI assistants to query screen history, compatible with Claude Desktop, Cursor, VS Code (Cline, Continue), and any MCP-compatible client. Zero-config command: claude mcp add screenpipe -- npx -y screenpipe-mcp@latest.
• Local REST API: Runs on localhost (default port 3030), provides endpoints for searching screen content, audio, and frames, allows raw SQL access to the underlying SQLite database, and provides a JavaScript/TypeScript SDK.

How to use

Download the macOS, Windows, or Linux desktop app, or run the CLI. After installation, you can perform local capture and search (works offline). Through MCP, you can let AI assistants query history, for example asking "what did I see in the past 5 minutes" or "summarize today's conversations." You can also create pipes to automate things, for example updating Linear every time task X is processed. If you use a coding agent to work on this repository, you must first read the screenpipe CLI skill, set up always-on local capture, verify capture freshness and storage, and then query history.

Use cases

• Knowledge workers who want to recall what they have seen and heard on their computer
• Developers who want to provide work context to AI coding assistants (Cursor, Claude Code, Cline, Continue)
• Researchers who need to search large amounts of screen information
• People with ADHD who frequently lose tabs, documents, and conversations
• Remote workers who want automatic meeting transcription and notes
• Teams and enterprises that want to deploy AI across the organization with deterministic data permissions and centralized configuration management
• Users who want a private, local-first alternative to cloud AI memory tools

Limitations and caveats

• Resource usage：CPU approximately 5-20% (varies with hardware, capture settings, transcription, and AI load); RAM approximately 0.5-3 GB for capture, with local models and additional load potentially requiring more; storage varies with activity, monitors, audio, capture settings, and retention policy, so it is recommended to measure a representative day first before planning storage. 8 GB of system memory is recommended. Linux capture behavior depends on the distribution, desktop environment, and permissions.
• Privacy and data paths: Screen frames, audio, transcriptions, and search indexes are stored locally on the device by default. Cloud AI, transcription, sync, integrations, exports, and enterprise storage may send data off the device. Local capture and search work offline once set up; cloud AI, sync, and connected services require a network. Local model inference such as Ollama is supported, and transcription, sync, integrations, and telemetry can be configured separately to control other network traffic. No account is required, and data can be exported, deleted, or backed up at any time. Optional end-to-end encrypted sync (zero-knowledge encryption).
• AI data permissions: Each pipe supports YAML frontmatter to control the data accessible to AI agents, including app and window filtering (allow-apps, deny-apps, deny-windows), content type controls (ocr, audio, input, accessibility), time and date restrictions, and endpoint gating (allow-raw-sql, allow-frames). It is enforced at three layers—skill gating, agent interception, and server middleware (per-pipe encrypted tokens)—and is deterministic rather than prompt-based.
• Licensing: Source builds are governed by LICENSE.md, which permits personal non-commercial use, non-profit/educational/research use, and limited organizational evaluation; commercial use requires a commercial license. Official builds are governed by the terms of service and any applicable subscription or app license. Existing lifetime licenses remain valid, but new lifetime purchases are no longer sold.
• Platforms and features: The desktop app has Free, Basic, and Business plans; please refer to the current pricing page for specific rates, monthly/annual billing, capacity options, and included features. Business includes personal device sync and team seat management, and currently does not include shared context between teammates. Enterprise pricing is determined by deployment scope.
• macOS audio exclusion: On macOS 14.4+, specificapplications can be excluded from system audio capture by listing their bundle IDs in ~/.screenpipe/audio-exclusions.json. Experimental CoreAudio System Audio must first be enabled in Settings → Recording. The exclusion list hot-reloads without requiring a restart. The "System Audio Recording Only" TCC permission is required.
• Enterprise features: Teams supports centralized configuration management, shared pipes, per-pipe AI data permissions, admin visibility, and override rules (employees can tighten but cannot loosen admin rules), and is MDM-ready (Intune, SCCM, Robopack, etc.). Enterprise deployments can offer SSO/SAML, audit logs, and SLA options. The security page states that the SOC 2 Type 2 report is available under NDA; SOC 2 is an attestation rather than a product certification. HIPAA applicability depends on the deployment, data flows, safeguards, and applicable agreements.

Integrations

Supports AI coding assistants (Cursor, Claude Code, Cline, Continue, OpenCode, Gemini CLI), AI chat assistants (ChatGPT via MCP, Claude Desktop via MCP, any MCP-compatible client), note-taking (Obsidian, Notion), local AI (Ollama, any OpenAI-compatible model server), and custom pipes automation.