Screenpipe

What mediar-ai/screenpipe MCP server does

mediar-ai/screenpipe MCP server gives AI agents access to a local activity record built from screen and audio capture. Screenpipe can capture accessibility data, screenshots, OCR fallback text, audio transcription, speaker information, keyboard inputs, and application changes. The resulting history is timestamped and stored on the local machine.

The system is intended for questions about recent or past computer activity, such as identifying what appeared on screen during a time window or summarizing conversations from the day. It also supports building agents, skills, and automations that respond to a user's activity. The repository describes SQL and embedding storage, semantic search, LLM-based history analysis, and event-triggered actions.

How it works

Screenpipe uses a recorder-first workflow: screen and audio capture run locally, the captured data is indexed, and an MCP process makes that history available to an agent. Screen capture uses the accessibility tree when available and falls back to OCR for sources such as remote desktops or games. Audio can include system output and microphone input, with transcription handled locally through Whisper or through Deepgram according to the documented configuration.

Captures are event-driven rather than based only on constant screenshots. App switches, clicks, typing pauses, scrolling, and other meaningful changes can trigger screen capture. This approach is described as reducing unnecessary processing and storage compared with recording identical frames continuously.

Setup and configuration

The repository provides a desktop application download and a CLI-based path. The CLI can start recording with npx screenpipe record, then configure the environment with npx screenpipe setup. The README also shows an MCP registration command using screenpipe-mcp:

claude mcp add screenpipe -- npx -y screenpipe-mcp@latest

After setup, an agent can ask questions such as what appeared in the last few minutes or request a summary of conversations from the current day. The setup command can install Screenpipe skills and MCP configuration into supported agents detected on the computer. The project documents macOS and Windows native installers and Linux support through building from source.

Screenpipe is designed for local operation and can work offline. Documented resource use is approximately 5–10% CPU, 0.5–3 GB RAM, and about 20 GB of storage per month in the main project description; another README section gives a lower 5–10 GB monthly disk estimate. Treat storage usage as workload-dependent. Filters can exclude windows, applications, Chrome extensions, passwords, and proprietary or sensitive data, and encryption at rest is optional.

Tools and capabilities

The mediar-ai/screenpipe MCP server is relevant for agents that need to:

• Search screen and audio history using natural-language queries.
• Retrieve recent activity for a time range.
• Summarize recorded conversations and computer work.
• Use indexed screen, transcription, and application context in agent workflows.
• Create pipes or automations that react to observed work.
• Access local history without sending the complete activity archive to a hosted service.

The README demonstrates asking an agent to create a pipe that updates Linear when work on a specified task is detected, but it does not provide the full pipe API or configuration details in the supplied material.

Limitations and notes

Continuous capture can consume substantial local resources and storage, particularly for long-running use. The system records sensitive computer activity, so capture filters, access controls, and local storage policies should be reviewed before deployment. Screenpipe is source-available under the Screenpipe Commercial License: personal, non-commercial use is permitted, while commercial use requires a license. The supplied material does not define the complete MCP tool schema, all configuration options, or the exact behavior of every supported agent client.