screenpipe

What it captures and how

Screenpipe is a tool that records your computer activity continuously and keeps that data local. The README describes it as a way to remember how you actually work, by capturing the screen and audio on the machine and then making that context available to agents such as Claude, Codex, Openclaw, Hermes, and Runner. The capture is explicit about being local first. The README states your data lives on your machine only, and that the code is source available so you can inspect, modify, and audit it. Technically the recorder captures the full accessibility tree, uses OCR as a fallback when the tree is not enough, and also handles transcription, speakers, keyboard inputs, and app switches. That combination means the history is not just pixels but structured events: which app was front, what was typed, and what was said. The README notes the resource cost, which is useful for setting expectations. Capture uses about 5 to 10 percent CPU, 0.5 to 3 GB of RAM, and roughly 20 GB of storage per month. Those numbers make clear this is a background service with a real footprint, not a lightweight occasional snapshot tool, and they help a user decide whether to run it on a laptop with limited disk.

Privacy controls and offline use

Because the tool watches everything on screen, the README spends real effort on privacy controls. There are filters that can be applied at the window level, the app level, through Chrome extensions, and with a passwords filter plus a proprietary AI PII model that aims to keep sensitive content out of the stored history. The data can be encrypted at rest if you enable that option. The recorder also works offline, so capturing does not require a cloud connection, and the local first design means the agent queries happen against the data on your machine. The README also describes a company brain mode where a team can share knowledge without turning the setup into surveillance, though the fine details of that sharing live in the linked enterprise page. For individual use, the practical privacy story is that you choose what to filter, you can encrypt the store, and nothing leaves the device unless you wire an agent to read it. The filters and the optional encryption are the levers a careful user should set before leaving continuous capture running, since once something is recorded it is in the local store subject to those rules. The README is candid that no automated filter is perfect, so the safest setting for sensitive work is to scope capture to specific apps and to keep encryption turned on.

Installing and connecting agents

There are two ways to run Screenpipe. The desktop app gives you every feature with auto updates, and the README links to a download page for it. The lighter path is the command line: you run npx screenpipe record to start capturing, then npx screenpipe setup to wire things up. The setup step can add the Screenpipe MCP server to supported agents, and the README shows the command for Claude as an example, after which you can ask questions like what you saw in the last five minutes or ask it to summarize today's conversations. There is also a CLI only setup for coding agents such as Claude Code, Codex, Gemini CLI, and Cursor, where the agent is given a skill file that covers a recorder first service default, an API only server mode, status checks in human and JSON form, local search, read only SQLite access, pipes, and connections. The pipes concept lets you build automations, such as updating a Linear task whenever you work on a specific item. The README points to docs, an enterprise page, and community channels for more. The Rust implementation fits the always on, low overhead goal, and the source available license marked NOASSERTION means the exact terms are in the repository's license file rather than a standard SPDX tag.

Editorial conclusion

The project is released under a license marked NOASSERTION and is written in Rust.