Screenpipe, a Y Combinator S26 (Summer 2026) batch company, launched on Hacker News via a Launch HN post on July 23, 2026. Founded by Louis, who began building the tool in 2024 as a CLI utility to record a user's screen and feed context to AI, Screenpipe records screen and audio locally on-device and gives AI agents a searchable memory of what the user has seen, said, and heard. The product was first posted to Hacker News by a community member in 2024, before its formal YC launch.

The capture pipeline is built around efficiency and structured data. Rather than continuously recording video and running OCR on every frame, Screenpipe listens for OS-level events such as app switches, clicks, typing pauses, and scrolling to detect meaningful changes. When a change is detected, it captures a screenshot paired with the operating system's accessibility tree at the same timestamp. OCR is used only as a fallback when structured accessibility data is unavailable. A February 2026 blog post for version 2.03 introduced accessibility APIs as the primary text-extraction method, claiming "100x less resource usage, 100% data accuracy" for that approach.

Audio is captured continuously, with speaker identification and transcription performed locally via Parakeet or Whisper, or optionally through cloud models. All indexed data is stored locally in a SQLite database, alongside mp4 files and, in some cases, markdown files.

For agent integration, Screenpipe exposes an AI-friendly API on port 3030, with authentication, MCP (Model Context Protocol) support, and a skills framework. The idea is that an AI agent can query a user's historical screen and audio context programmatically, retrieving what was on-screen or said in a meeting without the user manually surfacing that information.

Screenpipe also operates an extension ecosystem. A "pipe (app) store" within the platform distributes modular extensions called pipes, and the company has launched a developer program to encourage third-party contribution. A blog post referencing version 2.23, titled "screenpipe needs you," alluded to a YC launch squad, signaling ongoing team-building ahead of the formal batch launch.

The technical architecture is worth examining in the context of the broader agent ecosystem. MCP, originally introduced by Anthropic, has become a de facto standard for giving AI agents structured access to external tools and data sources. Screenpipe's choice to expose its API through MCP means that any MCP-compatible agent can, in principle, query a user's screen history as easily as it queries a file system or a database. The local-first storage model, combined with the accessibility-tree approach to text extraction, addresses two practical bottlenecks that have dogged ambient capture tools: resource overhead and data fidelity.

There are, of course, questions that the launch post does not fully resolve. Continuous screen and audio recording, even locally, sits in a category of tools that raises legitimate privacy and security considerations, particularly in enterprise or multi-tenant environments where a single machine may handle sensitive data from multiple contexts. Screenpipe stores everything on-device, which reduces the attack surface relative to cloud-dependent alternatives, but it also concentrates a comprehensive activity record on a single machine. How that data is protected at rest, and whether enterprises will permit such tools on managed devices, remains an open question that adoption, not architecture, will answer.

Looking at the trajectory from CLI tool to YC-backed company, the product has clearly matured. The shift from frame-by-frame OCR to event-driven accessibility-tree capture is a meaningful architectural improvement, and the MCP integration positions Screenpipe within the current agent-tooling stack rather than as a standalone utility. Whether the "pipe store" model will attract enough developer activity to create a durable ecosystem is the more speculative bet, and one that the developer program is clearly designed to test.

For the YC S26 batch, Screenpipe represents a specific bet: that AI agents need persistent, queryable memory of human activity to be genuinely useful, and that local-first capture is the mechanism that makes that palatable. The technology to build this has existed in pieces for some time; Screenpipe's contribution is the integration into a single, agent-accessible pipeline.