io.github.screenpipe/screenpipe-mcp

news

• 06/10 - we updated our license to keep screenpipe sustainable — more funding, more shipping, better product
• 05/29 - we released an alpha version of our AI PII model outperforming Google, Microsoft, and OpenAI models on computer recording data and running at 9ms on consumer device
• 05/14 - we joined YC S26

what is this?

screenpipe turns your computer into a personal AI that knows everything you've done. record. search. automate. all local, all private, all yours

┌─────────────────────────────────────────┐
│  screen + audio → local storage → ai   │
└─────────────────────────────────────────┘

• remember everything - never forget what you saw, heard, or did
• run agents that work based on what you do - pipes are agents triggered by your work activity

• search with ai - find anything using natural language
• 100% local - your data never leaves your machine
• source-available - inspect, modify, audit (LICENSE.md)

install

download the desktop app — all features, auto-updates

or run the CLI:

npx screenpipe record

then

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

then ask claude what did i see in the last 5 mins? or summarize today conversations or create a pipe that updates linear every time i work on task X

specs

• captures full accessibility tree, OCR as fallback, transcription, speakers, keyboard inputs, app switches
• 5-10% cpu usage
• 0.5-3gb ram
• ~20gb storage/month
• filters (window, app, chrome extensions, passwords, proprietary AI PII model)
• optional encryption at rest
• works offline

sdk

tauri, electron, swift sdk available today

Contributing

See CONTRIBUTING.md for guidelines, maintainers, and how to submit PRs. AI/vibe-coded PRs welcome!

Thanks to all contributors:

Translations

• 日本語 (Japanese)
• 简体中文 (Simplified Chinese)
• Français (French)
• Español (Spanish)
• Português BR (Brazilian Portuguese)
• Deutsch (German)
• Українська (Ukrainian)

Want to translate screenpipe into another language? See translated_readmes/.

About screenpipe

screenpipe is a source-available application that continuously captures your screen and audio, creating a searchable, AI-powered memory of everything you do on your computer. All data is stored locally on your device. It is the leading source-available alternative to Rewind.ai (now Limitless), Microsoft Recall, Granola, and Otter.ai. If you're looking for a rewind alternative, recall alternative, or a private local screen recorder with AI, screenpipe is the most popular option you can fully audit.

• Website: https://screenpi.pe
• Documentation: https://docs.screenpi.pe
• Discord: https://discord.gg/screenpipe
• License: Screenpipe Commercial License (source-available; personal, non-commercial use permitted, commercial use requires a license, see LICENSE.md)

Who screenpipe is for

• Knowledge workers who want to recall anything they've seen or heard on their computer
• Developers who want to give AI coding assistants (Cursor, Claude Code, Cline, Continue) context about what they're working on
• Researchers who need to search through large volumes of screen-based information
• People with ADHD who frequently lose track of tabs, documents, and conversations
• Remote workers who want automatic meeting transcription and notes
• Teams & enterprises who want to deploy AI across their organization with deterministic data permissions and central config management (screenpi.pe/team)
• Anyone who wants a private, local-first alternative to cloud-based AI memory tools

Platform support

Minimum requirements: 8 GB RAM recommended. ~5–10 GB disk space per month. CPU usage typically 5–10% on modern hardware thanks to event-driven capture.

Core features

Event-driven screen capture

Instead of recording every second, screenpipe listens for meaningful events — app switches, clicks, typing pauses, scrolling — and captures a screenshot only when something actually changes. Each capture pairs a screenshot with the accessibility tree (the structured text the OS already knows about: buttons, labels, text fields). If accessibility data isn't available (e.g. remote desktops, games), it falls back to OCR. This gives you maximum data quality with minimal CPU and storage — no more processing thousands of identical frames.

Audio transcription

Captures system audio (what you hear) and microphone input (what you say). Real-time speech-to-text using Whisper (Large-V3-Turbo) running locally on your device, or Deepgram for cloud transcription. Speaker identification and diarization. Works with any audio source — Zoom, Google Meet, Teams, or any other application.

On macOS 14.4+, you can exclude specific apps from system-audio capture by listing their bundle IDs in ~/.screenpipe/audio-exclusions.json. Enable Experimental CoreAudio System Audio in Settings → Recording first; the picker UI only appears once that flag is on.

{ "excluded_apps": [{ "bundle_id": "com.spotify.client", "name": "Spotify" }] }

The exclusion list hot-reloads — edits to the file and excluded apps launching/quitting are picked up on the engine's existing 500 ms tap-rebuild loop without restarting screenpipe. Override the file path with SCREENPIPE_AUDIO_EXCLUSIONS_PATH for testing. Note: this requires the "System Audio Recording Only" TCC permission in System Settings → Privacy & Security → Screen & System Audio Recording.

AI-powered search

Natural language search across accessibility-first screen text, OCR fallback text, and audio transcriptions. Filter by application name, window title, browser URL, date range. Full-text keyword search (SQLite FTS5) under the hood. Returns screenshots and audio clips alongside text results.

Timeline view

Visual timeline of your entire screen history. Scroll through your day like a DVR. Click any moment to see the full screenshot and extracted text. Play back audio from any time period.

Plugin system (Pipes)

Pipes are scheduled AI agents defined as markdown files. Each pipe is a pipe.md with a prompt and schedule — screenpipe runs an AI coding agent (like pi or claude-code) that queries your screen data, calls APIs, writes files, and takes actions. Built-in pipes include:

• meeting-summary: Summarizes the meeting that just ended and patches the note back onto the meeting record
• day-recap: Today's accomplishments, key moments, and unfinished work
• standup-update: What you did, what's next, and any blockers
• time-breakdown: Where your time went, by app, project, and category
• ai-prompt-journal: Captures every prompt you send to AI tools, saved to Obsidian or local markdown
• video-export: Create a video of your recent screen activity

Developers can create pipes by writing a markdown file in ~/.screenpipe/pipes/.

Pipe data permissions

Each pipe supports YAML frontmatter fields that give admins deterministic, OS-level control over what data AI agents can access:

• App & window filtering: allow-apps, deny-apps, deny-windows (glob patterns)
• Content type control: restrict to ocr, audio, input, or accessibility
• Time & day restrictions: e.g. time-range: 09:00-18:00, days: Mon,Tue,Wed,Thu,Fri
• Endpoint gating: allow-raw-sql: false, allow-frames: false

Enforced at three layers — skill gating (AI never learns denied endpoints), agent interception (blocked before execution), and server middleware (per-pipe cryptographic tokens). Not prompt-based. Deterministic.

MCP server (Model Context Protocol)

screenpipe runs as an MCP server, allowing AI assistants to query your screen history:

• Works with Claude Desktop, Cursor, VS Code (Cline, Continue), and any MCP-compatible client
• AI assistants can search your screen history, get recent context, and access meeting transcriptions
• Zero configuration: claude mcp add screenpipe -- npx -y screenpipe-mcp@latest

Developer API

Full REST API running on localhost (default port 3030). Endpoints for searching screen content, audio, frames. Raw SQL access to the underlying SQLite database. JavaScript/TypeScript SDK available.

Privacy and security

• 100% local by default: All data stored on your device in a local SQLite database. Nothing sent to external servers.
• Source-available: fully auditable codebase; personal, non-commercial use permitted.
• Local AI support: Use Ollama or any local model — no data sent to any cloud.
• No account required: Core application works without any sign-up.
• You own your data: Export, delete, or back up at any time.
• Optional encrypted sync: End-to-end encrypted sync between devices (zero-knowledge encryption).
• AI data permissions: Per-pipe YAML-based access control — deterministic enforcement at the OS level, not prompt-based. Three enforcement layers prevent AI agents from accessing unauthorized data.

How screenpipe compares to alternatives

Pricing

The source is available for personal, non-commercial use (see LICENSE.md). The signed desktop app uses a subscription:

• Standard: $25/month. Local-first capture, search, and timeline, all on your device.
• Pro: $50/seat/month. Everything in Standard plus cloud sync, cloud AI, and integrations. Teams buy 5+ seats self-serve.
• Enterprise: $150/seat/month. Managed deployment, central config, shared pipes, per-pipe AI data permissions, admin dashboard, SSO/SAML, MDM ready (Intune / SCCM). Sales-led. See screenpi.pe/team.

Existing lifetime licenses remain valid; new lifetime purchases are no longer sold.

Integrations

• 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
• Automation: Custom pipes (scheduled AI agents as markdown files)

Teams & enterprise

screenpipe Teams lets organizations deploy AI agents across their team with full control over what AI can access. See screenpi.pe/team.

• Central config management: Push capture settings (app filters, schedules, URL rules) to every device from an admin dashboard.
• Shared pipes: Deploy AI workflows (auto-standups, meeting-to-tickets, time tracking) team-wide.
• Per-pipe AI data permissions: YAML frontmatter controls what each pipe can access — apps, windows, content types, time ranges, endpoints. Enforced deterministically at the OS level via three layers (skill gating, agent interception, server middleware with per-pipe cryptographic tokens).
• Privacy boundary: Admins control what gets captured and what AI accesses. They never see the actual data — everything stays on each employee's device.
• Override rules: Employees can add stricter filters (e.g. also block personal email) but cannot weaken admin-set rules.
• MDM ready: Deploy via Intune, SCCM, Robopack, or any MDM solution.
• Enterprise: SSO/SAML, audit logs, SLA, SOC 2 / HIPAA compliance ready.

Technical architecture

1. Event-driven capture: Listens for OS events (app switch, click, typing pause, scroll, clipboard). When something meaningful happens, captures a screenshot + accessibility tree together with the same timestamp. Falls back to OCR when accessibility data isn't available. Idle fallback captures periodically when nothing is happening.
2. Audio processing: Whisper (local) or Deepgram (cloud) for speech-to-text. Speaker identification and diarization.
3. Storage: Local SQLite with FTS5 full-text search. Screenshots saved as JPEGs on disk (~300 MB/8hr vs ~2 GB with continuous recording).
4. API layer: REST API on localhost:3030. Search, frames, audio, elements, health, pipe management.
5. Plugin layer: Pipes — scheduled AI agents as markdown files. Agent executes prompts with access to screenpipe API.
6. UI layer: Desktop app built with Tauri (Rust + TypeScript).

API examples

Search screen content:

GET http://localhost:3030/search?q=meeting+notes&content_type=all&limit=10

Search audio transcriptions:

GET http://localhost:3030/search?q=budget+discussion&content_type=audio&limit=10

JavaScript SDK:

import { pipe } from "@screenpipe/js";

const results = await pipe.queryScreenpipe({
  q: "project deadline",
  contentType: "all",
  limit: 20,
  startTime: new Date(Date.now() - 24 * 60 * 60 * 1000).toISOString(),
});

Building from source

Check CONTRIBUTING.

Make sure to understand the main branch is moving fast and breaking things, if you're looking for a stable version check app releases https://github.com/screenpipe/screenpipe/releases and use the git commit accordingly (production app is behind paywall).

Frequently asked questions

How much does screenpipe cost? The signed desktop app uses a subscription starting at $25/month; existing lifetime licenses remain valid. The source is available for personal, non-commercial use, so you can build and run it yourself (see LICENSE.md); commercial use of the source requires a license.

Does screenpipe send my data to the cloud? No. All data is stored locally by default. You can use fully local AI models via Ollama for complete privacy.

How much disk space does it use? ~5–10 GB per month. Event-driven capture only stores frames when something changes, dramatically reducing storage compared to continuous recording.

Does it slow down my computer? Typical CPU usage is 5–10% on modern hardware. Event-driven capture only processes frames when something changes, and accessibility tree extraction is much lighter than OCR.

Can I use it with ChatGPT/Claude/Cursor? Yes. screenpipe runs as an MCP server, allowing Claude Desktop, Cursor, and other AI assistants to directly query your screen history.

Can it record multiple monitors? Yes. screenpipe captures all connected monitors simultaneously.

How does text extraction work? screenpipe primarily uses the OS accessibility tree to get structured text (buttons, labels, text fields) — this is faster and more accurate than OCR. When accessibility data isn't available (remote desktops, games, some Linux apps), it falls back to OCR: Apple Vision on macOS, Windows native OCR, or Tesseract on Linux.

Can I deploy screenpipe to my team? Yes. Screenpipe Teams provides central config management, shared AI pipes, and per-pipe data permissions. Admins control what gets captured and what AI can access — employees' actual data never leaves their devices. See screenpi.pe/team.

How do AI data permissions work? Each pipe supports YAML frontmatter fields (allow-apps, deny-apps, deny-windows, allow-content-types, time-range, days, allow-raw-sql, allow-frames) that deterministically control what data the AI agent can access. Enforcement happens at three OS-level layers — not by prompting the AI to behave. Even a compromised agent cannot access denied data.

Company

Built by screenpipe (Mediar, Inc.). Founded 2024. Based in San Francisco, CA.

• Founder: Louis Beaumont (@louis030195)
• Twitter: @screenpipe
• Email: louis@screenpi.pe