screenpipe: local screen recording as AI memory
/screenpipe
screenpipe is a Rust-based desktop app that records your screen and audio locally, all day, and turns that into searchable memory an AI agent can query. Reach for it if you want Claude, Cursor, or an MCP client to know what you actually did today without re-explaining it. Skip it if $25/month for the signed app, or self-building from source, feels like too much for a screen recorder.
What is screenpipe?
screenpipe is a source-available desktop application that captures your screen and audio 24/7 and stores everything in a local SQLite database you can search with natural language. Each capture pairs a screenshot with the OS accessibility tree instead of relying only on OCR, and the whole history is exposed through a REST API and an MCP server so tools like Claude Desktop or Cursor can query what you've been working on.
Core Capabilities
- ✓Watches for OS-level signals — switching apps, clicking, pausing mid-type, scrolling — and only fires a screenshot when one of those actually happens, instead of grabbing a frame on a timer. Falls back to OCR (Apple Vision, Windows OCR, Tesseract) only when the accessibility tree isn't available, like over remote desktop or in a game.
- ✓Local speech-to-text with Whisper (Large-V3-Turbo) running on your device, with an option to switch to Deepgram for cloud transcription; includes speaker identification and diarization for both system audio and microphone input.
- ✓Full-text search (SQLite FTS5) across screen text, OCR fallback text, and audio transcripts, filterable by app name, window title, browser URL, and date range.
- ✓Pipes — markdown files with a prompt and a schedule that let an AI coding agent (pi, claude-code) query your captured history and act on it; built-ins include meeting-summary, day-recap, standup-update, and time-breakdown.
- ✓Per-pipe YAML permissions (allow-apps, deny-apps, time-range, allow-raw-sql, allow-frames) enforced at three layers — skill gating, agent interception, and server middleware with per-pipe tokens — so a compromised agent still can't reach denied data.
- ✓MCP server support for Claude Desktop, Cursor, and VS Code (Cline, Continue) with a one-line setup: `claude mcp add screenpipe -- npx -y screenpipe-mcp@latest`.
- ✓A REST API on localhost:3030 with raw SQL access to the underlying SQLite store, plus a JavaScript/TypeScript SDK for building your own tools on top of the captured data.
Who Benefits from screenpipe?
- •Developers who want Cursor, Claude Code, Cline, or Continue to have context on what they were actually doing, not just what's in the current file.
- •Knowledge workers and researchers who need to find something they saw or heard earlier without remembering which app or document it was in.
- •People with ADHD who lose track of tabs and conversations and need a way to reconstruct what happened.
- •Remote workers who want meeting transcription and notes generated automatically from system audio.
- •Teams and enterprises that need central config and per-pipe data permissions before they'll let AI touch employee screens.
Getting Started with screenpipe
The desktop app is the easiest path — download the signed .dmg (macOS Intel and Apple Silicon) or .exe (Windows 10/11) installer, which auto-updates itself. Linux has full support too, but only by building from source, per the README's platform table. If you'd rather skip the installer, run the CLI with `npx screenpipe record`, then either `npx screenpipe setup` or wire it straight into Claude with `claude mcp add screenpipe -- npx -y screenpipe-mcp@latest`. After that you can ask Claude things like "what did I see in the last 5 mins?" directly. One note: the README says the main branch moves fast and can break, so pin to a tagged release if you want stability — the production app itself sits behind the paywall.
Advantages of screenpipe's Approach
- ✓Reads the OS accessibility tree first and only drops to OCR when it has to, so text extraction comes out faster and more accurate than tools that OCR every frame.
- ✓Everything — screenshots, audio, transcripts, the search index — lives in a local SQLite database, and the core app runs with no account required.
- ✓Per-pipe YAML permissions give admins deterministic, OS-level control over what an AI agent can touch, enforced at three separate layers rather than by asking the model nicely.
- ✓Captures every connected monitor, not just the active window, and exposes a REST API plus JS/TS SDK for building on top of the captured data.
- ✓The pipe system turns captured activity straight into automation — meeting-summary and day-recap ship as built-ins you can adapt instead of writing from scratch.
Considerations Before Using
- △The signed desktop app is a subscription — $25/month for Standard, $50/seat/month if you want cloud sync or cloud AI; the source itself is free only for personal, non-commercial use.
- △Telemetry (PostHog analytics, Sentry crash reports) is on by default in the desktop app — you have to go into Settings → Privacy → Analytics to turn it off, and cloud transcription or hosted AI still sends that specific data out.
- △Linux has no native installer — you build from source, and the README warns the main branch moves fast and breaks things, so you're expected to track a tagged release.
- △Storage adds up fast — the README puts it at roughly 5-10 GB per month with event-driven capture, and that's before you factor in a busy multi-monitor setup.
- △Running a 24/7 record of your screen and audio is a lot of surface area to get wrong — a misconfigured pipe permission or an exposed sync key leaks more than a typical app would.
Comparing with Other Memory Tools
Common Questions
The signed desktop app runs $25/month for Standard, $50/seat/month for Pro (adds cloud sync and cloud AI), and $150/seat/month for Enterprise with central admin controls. Existing lifetime licenses still work, but screenpipe no longer sells new ones. The source is free to build and run yourself for personal, non-commercial use.
Screen frames, audio, transcripts, and the search index stay local by default. That said, the desktop app ships with PostHog analytics and Sentry crash reporting on out of the box, and choosing cloud transcription, hosted AI, or cloud sync sends that specific data out — you can turn telemetry off in Settings → Privacy → Analytics.
screenpipe's platform-support notes and FAQ put it at roughly 5–10GB per month, though the specs section elsewhere in the same README cites closer to 20GB. Either way, event-driven capture — only storing frames when something actually changes — keeps it well below continuous recording.
Yes — screenpipe captures every connected monitor at once, not just the active window, which the README calls out as a difference from Rewind/Limitless (active window only) and Granola (meetings only).
Each pipe carries YAML frontmatter — allow-apps, deny-apps, deny-windows, time ranges, and switches like allow-raw-sql or allow-frames — that screenpipe enforces at three separate layers: skill gating, agent interception, and server middleware with per-pipe cryptographic tokens. It's deterministic access control, not a prompt asking the model to behave.
Yes, through Screenpipe Teams — it adds central config management, shared pipes, and per-pipe data permissions from an admin dashboard, while each employee's actual captured data stays on their own device. Enterprise adds SSO/SAML and MDM support for Intune or SCCM.
The problem it solves
Screen recorders capture pixels. Note apps capture only what you typed by hand. Neither gives an AI agent the actual sequence of apps, clicks, and conversations that led to a decision, so tools like Cursor or Claude Code end up working from whatever you paste into the chat box. screenpipe's answer is to record the accessibility tree and audio continuously and hand agents a queryable history instead of a summary you wrote after the fact.
Who should try it — and who should skip
Reach for screenpipe if you're already scripting agents against Claude Code, Cursor, or another MCP client and want them to see what you actually did instead of what you remembered to paste in — local-only storage and the per-pipe permission system make that workable even with sensitive screen content. Skip it if you just want occasional screenshots or a simple meeting-transcript app; the always-on capture, $25+/month signed app, and default telemetry are more machinery than that job needs. Enterprises that need audited, permissioned AI access to employee screens are the other clear fit, given the SOC 2/HIPAA-ready tier and deterministic pipe permissions.
Related repositories
Still deciding about screenpipe?
One click hands the question to an AI along with this page — see what it says about screenpipe.