Screenpipe, a company in the Y Combinator Summer 2026 batch, launched on Hacker News on July 23, 2026. Founded by Louis, who started building the tool in 2024, Screenpipe records your screen and audio on your own computer and gives AI assistants a searchable record of what you have seen, said, and heard. A community member first posted the tool to Hacker News in 2024, before its formal YC launch.

The tool is designed to be efficient. Instead of recording video constantly and scanning every frame for text, Screenpipe watches for moments when something meaningful changes, such as when you switch apps, click, pause while typing, or scroll. When it detects a change, it takes a screenshot and also grabs a structured description of what is on screen from the operating system. Think of it like a screen reader that narrates a page for a visually impaired user: instead of guessing text from pixels, it reads the underlying structure directly. A February 2026 blog post for version 2.03 introduced this approach, claiming "100x less resource usage, 100% data accuracy."

Audio is captured continuously, with transcription and speaker identification performed on your device using tools called Parakeet or Whisper, or optionally through cloud models. Everything is stored locally on your computer in a database, alongside video files and, in some cases, markdown files.

For AI integration, Screenpipe provides an API that lets an AI assistant look up what was on your screen or said in a meeting without you having to find and share that information manually. It supports a standard called MCP, short for Model Context Protocol, which was introduced by the AI company Anthropic and has become a common way for AI assistants to connect to external tools and data.

Screenpipe also has an extension system. A built-in "pipe store" distributes small add-on modules called pipes, and the company has launched a developer program to encourage others to build them. A blog post for version 2.23, titled "screenpipe needs you," hinted at a YC launch squad, signaling the team was still growing before the formal batch launch.

The broader context is how Screenpipe fits into the growing world of AI assistants. Because it uses MCP, any compatible AI assistant can in principle search your screen history as easily as it searches a folder of documents. Storing everything on your own computer, rather than in the cloud, and using the operating system's built-in text data instead of scanning pixels, solves two problems that have held back similar tools: they tend to drain your computer's resources and sometimes get the text wrong.

There are questions the launch does not fully answer. Recording your screen and audio continuously, even locally, raises real privacy and security concerns, especially in workplaces where one computer may handle sensitive information from different people or projects. Storing everything on-device means there is no cloud server to hack, but it also means a complete record of your activity lives on one machine. How that data is protected, and whether companies will allow such tools on managed devices, is a question that adoption will answer over time.

Looking at the path from a simple command-line tool to a YC-backed company, the product has clearly grown up. The switch from scanning every frame for text to using the operating system's structured data is a meaningful improvement, and the MCP support places Screenpipe within the current ecosystem of AI tools rather than as a standalone app. Whether the "pipe store" will attract enough developers to build a lasting ecosystem is the more speculative bet, and one that the developer program is designed to test.

For the YC S26 batch, Screenpipe is making a specific bet: that AI assistants need a persistent, searchable memory of what you do to be genuinely useful, and that keeping all of that data on your own computer is what makes it acceptable. The technology to do this has existed in pieces for some time; Screenpipe's contribution is putting it together into one pipeline that an AI assistant can actually use.