Screenpipe

What Screenpipe does

Screenpipe provides local-first workflow memory for AI agents by continuously recording and indexing desktop screen and audio activity on your computer. Through its Model Context Protocol (MCP) server, AI assistants--such as Claude Desktop, Cursor, Windsurf, and Claude Code--can retrieve timestamped context, search past work, summarize conversations, and generate automated workflows grounded in actual user activity.

How it works

Screenpipe operates as an on-device daemon that indexes desktop activity locally:

• Screen capture: Extracts text and UI state via the accessibility tree where available, falling back to OCR when needed.
• Audio capture: Records system and microphone audio, providing local transcription via Whisper or optional cloud transcription.
• Local storage: All captured frames, transcripts, and embeddings reside in a local SQLite database on your machine by default.
• MCP server bridge: The screenpipe-mcp package runs locally to safely answer queries from connected AI agents without exposing the raw database directly.

Installation & Setup

Recommended: Screenpipe Desktop Application

The most reliable and recommended installation path is via the desktop application:

1. Install the desktop app from screenpipe.com.
2. Open Settings > Connections (or configure during onboarding).
3. Connect Claude Desktop.

This automatically configures Claude Desktop using Screenpipe's bundled runtime and injects your SCREENPIPE_LOCAL_API_KEY directly into the client configuration with zero external Node or PATH dependencies.

Manual NPX Configuration

If you prefer configuring MCP manually or are connecting clients like Cursor or Windsurf, configure your client's MCP settings:

{
  "mcpServers": {
    "screenpipe": {
      "command": "npx",
      "args": ["-y", "screenpipe-mcp@latest"],
      "env": {
        "SCREENPIPE_LOCAL_API_KEY": "YOUR_LOCAL_API_KEY"
      }
    }
  }
}

You can obtain your local API key anytime by running:

screenpipe auth token

Available Tools

The standard stdio MCP package provides six core tools:

• search_content: Search captured screen recordings, OCR text, and audio transcriptions over arbitrary time windows.
• export-video: Export video clips from captured screen history.
• list-meetings: Inspect and retrieve detected meeting events and transcripts.
• activity-summary: Generate structured summaries of computer activity over a designated timeframe.
• search-elements: Search and inspect captured UI elements and accessibility tree items.
• frame-context: Retrieve frame context and OCR details for specific timestamps.

Orgs with enterprise query gateways can additionally configure SCREENPIPE_ENTERPRISE_TOKEN and SCREENPIPE_TEAM_API_URL to enable team-* org-wide search tools.

Privacy, Security, and Licensing

• Privacy by default: Screenpipe stores all recordings, transcriptions, and indexes locally on your computer.
• Cloud transmission: Cloud AI, remote transcription (e.g. Deepgram), sync, and external integrations transmit data only if explicitly configured.
• Model provider context: When an AI assistant invokes Screenpipe tools, the retrieved snippet context is transmitted to the assistant's selected model provider as part of the prompt conversation.
• License: Screenpipe is source-available under the Screenpipe Commercial License (not OSI open source). Personal and non-commercial usage is free; commercial use requires a commercial license.