Launch HN: Screenpipe (YC S26) – Record how you work and turn that into agents

Louis, the founder behind Screenpipe, has spent years chasing the same itch: he wants an AI that actually knows what he did today, not just what he typed into a chat box. He's been keeping a personal "second brain" since 2020, tinkering with GPT-2 fine-tunes and early RAG setups, and built one of the first Obsidian AI plugins before that project evolved into an API called Embedbase. Screenpipe is the next iteration of that obsession, and it launched on Hacker News this week as part of YC's Summer 2026 batch.

The pitch is straightforward even if the engineering underneath isn't. Screenpipe watches your screen and listens to your mic, entirely on your own machine, and builds a searchable record of what you saw, said, and heard throughout the day. Feed that into an agent and, in theory, you can ask it to summarize what you accomplished between 8am and 4pm, or have it quietly update your CRM every time you visit a LinkedIn profile. The goal isn't a smarter chatbot. It's giving AI the same working context a human colleague would have, without you having to explain your job to it every single time.

Louis started the project as a personal CLI tool in 2024, and an early HN post about it shaped a lot of what came next. The original version just recorded video continuously and ran OCR over every frame, which sounds reasonable until you realize it turns a laptop into a space heater and produces mountains of duplicate junk. The current build is smarter about it: Screenpipe listens for actual events, app switches, clicks, typing pauses, idle gaps, and only then grabs a screenshot paired with the operating system's accessibility tree, falling back to OCR only when that structured data isn't available. Audio gets transcribed locally through Parakeet or Whisper, with speaker identification baked in.

All of it lands in a local SQLite database alongside mp4 and markdown files, and an API on port 3030 lets agents like Claude, ChatGPT, or Screenpipe's own chat interface query that history directly, with MCP and skill support included. For companies worried about leaking sensitive data, the team built a custom PII redaction model that runs locally via Apple's MLX or Windows DirectML, with a cloud fallback for underpowered machines. There's also an enterprise tier aimed at surfacing automation opportunities across a whole team's workflow, though in that case the company gets to decide where its own data lives.

What's notable is where Louis places Screenpipe in the arc of AI memory tools. He walks through the whole lineage, fine-tuning, tool calling, MCP, agent skills, Karpathy-style LLM-maintained wikis, and argues all of them still require someone to manually feed the AI its sources. Screenpipe's bet is that the next real unlock isn't a better memory format. It's giving AI eyes and ears on the actual work happening across every app on your desktop, unprompted.