Go back to Learn

Tutorial · AI agents · Automation · About 10 min read

screenpipe: give Claude, Codex and Cursor a memory of your screen

Install screenpipe, connect its local screen and meeting history to your AI agent via MCP, and understand the privacy settings and real costs.

Start tutorial
A screen timeline feeds context to the screenpipe mark on a dark board.
Paulo Cauhi · Last verified

Prefer a PDF? Get it by email.

You get the PDF and future guides by email. Unsubscribe any time. Privacy

Your agent only knows what you paste into it. screenpipe keeps a searchable, local history of your screen text, app activity and transcribed calls. Its MCP server gives relevant excerpts to Claude, Codex and Cursor when you ask. It needs operating-system permissions, enough storage and a privacy check first.

As of 30/09/2026, founder/CEO Louis Beaumont's company was in Y Combinator S26 (YC also backed Airbnb and Dropbox); the repository had 21,781 stars. screenpipe reads the OS accessibility tree, falls back to OCR, records app switches and system/microphone audio, and lets you search by app, window, URL or date. It transcribes Zoom, Meet and Teams without a meeting bot.

Official screenpipe app screenshot showing the timeline and app icons.
Real screenshot from the project README, not your own history.

Install screenpipe

Ready-made app: Download at screenpipe.com. On 30/09/2026 the Free tier covered one device and limited searchable history; Basic was US$25/month or US$250/year for full history; Business US$50/seat/month; Enterprise custom. Check current prices. The app includes Timeline, Meetings and settings.

Free headless CLI for permitted personal use: Install Node.js 18+ and run:

Terminal
npx -y screenpipe@latest record

Leave the terminal running. This records locally and exposes the API on http://localhost:3030, without the app UI. Check privacy before recording a shared machine.

Build the app from source: Follow the current CONTRIBUTING.md. On macOS install rustup, Homebrew, Bun and full Xcode, then:

Terminal
brew install pkg-config ffmpeg jq cmake wget git-lfs
Terminal
git clone https://github.com/screenpipe/screenpipe.git && cd screenpipe
Terminal
cargo build --release --features metal
Terminal
cd apps/screenpipe-app-tauri && bun install && bun run build:tauri:dev

This is a development build, not a packaged installer. Windows needs VS Build Tools, Rustup, LLVM, CMake, Git, 7zip and Bun via winget; Linux users should follow the source-build instructions. Commercial use of source-built code requires a paid licence.

Set permissions and privacy

On macOS open System Settings → Privacy & Security and grant screenpipe Screen & System Audio Recording, Microphone and Accessibility; restart the app if prompted. In Settings → Recording, select local Whisper instead of cloud transcription. In Settings → AI, select a local Ollama model instead of screenpipe-cloud, with Ollama installed and running. The project's own screenshots show cloud options selected; do not assume offline operation out of the box.

Official screenpipe recording settings showing the transcription engine.
Check the selected transcription engine in your version.

Recordings, transcripts and index stay on your device by default, with optional encryption at rest. PostHog analytics and Sentry crash reports are on by default; turn them off at Settings → Privacy → Analytics. Cloud transcription, screenpipe cloud AI and sync transfer data when chosen. Sending retrieved excerpts through MCP to Claude or Codex sends them to Anthropic or OpenAI; Cursor may also use a cloud model. For fully local work, use local capture, Whisper and Ollama in screenpipe without external AI clients. See data flow and the FAQ.

Official screenpipe privacy settings with encryption and ignored-window options.
Review analytics, encryption and excluded apps.

Password fields are skipped by default (skip_password_fields); form values such as passwords, cards and secrets are redacted on-device by default. 1Password, Bitwarden, LastPass and Dashlane are excluded. The filter docs warn redaction can miss information. Add banking and sensitive apps to ignored windows; inspect what was captured before sharing.

Connect Claude Code

Keep screenpipe running at localhost:3030; Node 18+ is needed. Run:

Terminal
claude mcp add screenpipe --transport stdio --scope user -- npx -y screenpipe-mcp

Open Claude Code and type /mcp to check the server. Approve search tools when asked. Restart Claude Code and check the local service if disconnected.

Connect Claude Desktop

Go to Settings → Connections and install the screenpipe extension. Alternatively, merge this server into claude_desktop_config.json, save and restart Desktop. Preserve existing servers:

{"mcpServers":{"screenpipe":{"command":"npx","args":["-y","screenpipe-mcp"],"transport":"stdio"}}}

Check that screenpipe's tools appear in chat.

Connect Codex

In the Codex app use Settings → Connections → Connect (writes ~/.codex/config.toml). Or append to that file without removing other servers:

[mcp_servers.screenpipe]
command = "npx"
args = ["-y", "screenpipe-mcp"]
enabled = true

Restart Codex and confirm the connection. Retrieved excerpts reach OpenAI if your Codex model runs in its cloud.

Connect Cursor and ask a question

Use the Cursor deep link in the MCP guide, or Cursor Settings → MCP → Add global server, command npx, arguments -y and screenpipe-mcp. For project-scoped setup, merge into .cursor/mcp.json:

{"mcpServers":{"screenpipe":{"command":"npx","args":["-y","screenpipe-mcp"]}}}
Official Cursor screenshot with screenpipe MCP audio search results.
Published MCP result, not a guarantee about your own recordings.

Shortcut: npx screenpipe setup installs skills and MCP for supported agents it detects. Prefer manual setup if you want to restrict access. Ask “what did I see on screen in the last 5 minutes?”, “summarize today's conversations” or “what price did the client quote on yesterday's call?” The price question is illustrative: it requires a recorded, legible transcript. Verify answers in the timeline.

Meetings, storage and licence

The Summarize button works on demand. When enabled, the bundled meeting-summary pipe responds to meeting_ended and writes a summary into the meeting note. Its default preset is screenpipe-cloud, with paid-plan AI allowance; automatic summaries are not guaranteed unless configured. Under Pipes → meeting-summary, choose your own local Ollama model where available, enable it and test a meeting. See meeting intelligence.

As of 30/09/2026 the vendor README claimed 5–10% CPU, 0.5–3 GB RAM and ~20 GB/month storage. Its pricing page instead listed 10–30% CPU on Apple M3, 600 MB–4 GB RAM and ~30 GB/month. These are not independent benchmarks. Set retention and delete old data under Settings → Storage.

Official screenpipe Storage settings showing disk and retention controls.
Review retention and disk usage before long recordings.

Since 10/06/2026 current source is source-available, not open source, under the Screenpipe Commercial License. Personal non-commercial, nonprofit, education and research use are free; companies get seven days to evaluate, then need a paid licence. Older MIT versions retain their old licence. The getting-started page still says MIT and is stale. Official binaries have separate subscription terms.

FAQ

Is it free? The ready-made Free plan has limited history; CLI/source use is free only where the licence permits. Basic was US$25/month on 30/09/2026.

Does my screen go to the cloud? Storage is local by default, but telemetry is enabled; cloud options and context shared with remote AI clients leave your machine.

Does it record passwords? Password fields are skipped and form values redacted by default, but filtering may miss information. Exclude sensitive apps.

Will it slow down my Mac? It depends. Check actual performance on your machine rather than trusting the vendor's differing estimates.

Does it work on Windows/Linux? Yes, with different permissions and installation steps. For Linux builds consult CONTRIBUTING.md.

Can I use it at work? Source-built commercial use needs a paid licence; also check your workplace's recording and consent policies.

How do I delete history? Use Settings → Storage to remove old data and set retention. Uninstalling the app alone may leave data files.

Keep learning

Choose a practical next step.

Go back to Learn

Get the guide as a PDF

Get this guide as a PDF, plus future practical guides by email.

You get the PDF and future guides by email. Unsubscribe any time. Privacy

Comments

Have a question or something to add? Join the conversation.

No account needed. Select Sign in and choose a name to comment.

Loading the conversation…

Comment privacyCommenting guidelines