Screenpipe
Record your screen 24/7 locally and turn real work into searchable memory for your AI agents.
Evidence shows: README clearly states data is stored locally by default, offers optional encrypted sync and local AI support, and discloses telemetry (PostHog, Sentry) and cloud feature data flows. Pipes have YAML-based permission controls, but no user confirmation mechanism is provided. Dependency management uses pinned versions and git references, but no vulnerability scanning evidence is provided. External effects include network requests (telemetry, cloud features), but no rollback mechanism is provided. Source attribution: license and author info are clear, but publisher is unverified. Deductions: user confirmation mechanism missing, dependency security evidence insufficient, rollback mechanism not mentioned.
Evidence shows: README and test files describe a consistent architecture and features, but no execution results are provided. Dependency availability: pinned versions and git references are used, but no availability guarantees are provided. Failure messages: test files check error statuses, but no user-visible error messages are provided. Deductions: dependency availability evidence insufficient, failure messages not explicit.
Evidence shows: README describes multiple audiences (knowledge workers, developers, researchers, etc.) and scenarios (meeting notes, search, automation). Capability boundaries: Pipes have permission controls, but not all limitations are explicit. Trigger precision: Pipes have schedules and permissions, but no precise trigger conditions are provided. Environment fit: multi-platform support is mentioned, but no system requirement details are provided. Deductions: trigger precision evidence insufficient, environment fit details limited.
Evidence shows: README provides detailed information architecture, including installation, usage, API examples, and FAQ. Install notes are clear, but no detailed system requirements are provided. Naming stability: version numbers exist, but no naming conventions are provided. Examples and FAQ are rich. Known limitations: not explicitly listed. License: clearly a commercial license, but metadata is NOASSERTION. Versioning: version numbers exist, but no changelog is provided. Maintenance responsibility: author and contact info are present, but no maintenance policy is explicit. Deductions: naming stability, known limitations, versioning evidence insufficient.
Evidence shows: Output usability: API and SDK provide structured output. Marginal value: unique features compared to alternatives (local, multi-platform, plugin system). Cost-benefit: pricing is clear, but no performance benchmarks are provided. Deductions: cost-benefit evidence insufficient.
Evidence shows: Claims in README are partially supported by documentation, but no independent verification is provided. Cross-source corroboration: no external verification is provided. Fact-inference separation: not explicitly distinguished. Deductions: lack of independent verification and fact-inference separation.
- Publisher identity is unverified; proceed with caution.
- Dependency security lacks vulnerability scanning evidence; review dependencies.
- Telemetry is enabled by default; be mindful of privacy settings.
- License is commercial; commercial use requires payment.
What does this agent do, and when should you use it?
screenpipe is a source-available desktop application that continuously captures your screen and audio, turning what you see, hear, and do into a searchable local memory. It features event-driven screen capture, local speech-to-text (Whisper), AI-powered search, a timeline view, and a plugin system called Pipes for scheduling AI agents expressed as markdown files. It runs as an MCP server, allowing AI assistants like Claude Desktop and Cursor to query your screen history directly. All data is stored locally on your device by default, with optional encrypted sync, and the code is fully auditable. The app targets macOS, Windows, and Linux; the signed desktop app is subscription-based, while the source is available for personal non-commercial use.
screenpipe listens for OS events (app switches, clicks, typing pauses, scrolling) and captures screenshots only when meaningful changes occur, pairing each capture with the accessibility tree and falling back to OCR when accessibility data is unavailable. It transcribes system audio and microphone input using local Whisper (Large-V3-Turbo) or optional Deepgram cloud transcription, with speaker identification. Captured data is stored in a local SQLite database with FTS5 full-text search. It exposes a REST API on localhost:3030 for searching screen content, frames, audio, elements, and health. The Pipes system schedules AI agents—markdown files with prompts and schedules—executed by coding agents like pi or claude-code, with built-in pipes for meeting summaries, day recaps, standup updates, time breakdowns, and AI prompt journals. It also provides an MCP server (screenpipe-mcp) for integration with AI coding assistants such as Cursor, Claude Code, and Cline.
- Knowledge workers who need to recall something they saw or heard on their computer days ago, such as a specific document or meeting discussion.
- Developers who want to give AI coding assistants like Cursor or Claude Code context about what they're working on, including recent code and relevant web pages.
- Remote workers who want automatic meeting transcription and notes without manual effort.
- People with ADHD who frequently lose track of tabs, documents, and conversations and need a searchable memory.
- Teams that want to deploy AI workflows like automatic standup updates or time tracking while maintaining strict control over what AI can access.
- Researchers who need to search through large volumes of screen-based information such as papers, chat logs, and code snippets.
What are this agent's strengths and limitations?
- Event-driven capture keeps CPU usage at 5–10% and storage at ~5–10 GB/month, far less than continuous recording.
- 100% local by default, with support for local AI models (Ollama), no account required, and fully auditable source code.
- Robust Pipes system with deterministic per-pipe YAML data permissions enforced at three layers (skill gating, agent interception, server middleware with cryptographic tokens).
- Cross-platform support for macOS, Windows, and Linux, and captures all monitors simultaneously.
- Seamless integration with major AI assistants via MCP, including Claude Desktop and Cursor.
- The full desktop app requires a subscription (Standard $25/mo, Pro $50/seat/mo, Enterprise $150/seat/mo); the source is limited to personal non-commercial use, commercial use needs a license.
- Resource-heavy: RAM usage 0.5–3 GB and storage ~20 GB/month, which may be prohibitive on lower-end hardware.
- Platform support is uneven: Linux requires building from source, and some features like system audio capture may depend on OS permissions.
- Product analytics (PostHog) and crash reporting (Sentry) are enabled by default, requiring manual opt-out for privacy-conscious users.
- The source is available, but the production-grade app is behind a paywall; self-hosting may not deliver the same features or stability.
How do you install or deploy this agent?
Download the desktop app at https://screenpipe.com/how-to-install?download=1 for a native installer with auto-updates. Alternatively, run the CLI: npx screenpipe record to start recording, then npx screenpipe setup to configure. For MCP integration, run claude mcp add screenpipe -- npx -y screenpipe-mcp@latest.
How do you use this agent?
After installation, the app records automatically in the background. Search your screen history using natural language, e.g., ask Claude Desktop (via MCP): "what did I see in the last 5 mins?" or "summarize today conversations". To create custom automations, write a markdown file with a prompt and schedule in ~/.screenpipe/pipes/. Use the REST API for programmatic access: GET http://localhost:3030/search?q=meeting+notes&content_type=all&limit=10 to search screen content.
How does this agent compare with similar options?
The README compares screenpipe to Rewind.ai (Limitless), Microsoft Recall, and Granola. It claims advantages such as source-available code, multi-platform support (macOS, Windows, Linux), 100% local data storage, multi-monitor capture, local Whisper audio transcription, a full REST API, and a plugin system, which competitors often lack. It also positions itself as a leading alternative to Rewind and Recall.
FAQ
Does screenpipe send my data to the cloud?
Is screenpipe free to use?
Can I record multiple monitors?
How do AI data permissions work?
allow-apps, deny-apps, deny-windows, and allow-content-types to deterministically control what data the AI agent can access. Enforcement happens at three OS-level layers—skill gating, agent interception, and server middleware with per-pipe cryptographic tokens—not by prompting the AI to behave. Even a compromised agent cannot access denied data.