Louis Beaumont has a straightforward pitch, even if the implications aren't: let software record everything you do on your computer, but keep the data on your machine. No cloud uploads. No Microsoft servers. Just you, your local SQLite database, and whatever AI agents you decide to unleash on eight hours of screenshots and transcribed audio.
It's the kind of idea that makes privacy advocates wince and productivity obsessives lean forward.
Screenpipe, the six-person startup Beaumont founded, emerged from Y Combinator's Summer 2026 batch with a tool that fills a peculiar gap in the market—one created partly by Microsoft's stumbles and partly by the growing appetite for AI assistants that actually know what you've been working on. The timing feels almost too convenient. Microsoft's Recall feature, announced to considerable fanfare in mid-2024, promptly triggered a backlash over security concerns and has languished in development purgatory ever since. Limitless, the well-funded outfit previously known as Rewind AI, got acquired by Meta in December 2025, leaving its product roadmap in limbo.
Which leaves Beaumont arguing that local control was the obvious answer all along. Whether developers and enterprises agree is another question entirely.
The Mechanics of Always-On Capture
Here's what Screenpipe actually does. The software takes screenshots and audio clips as you work—not continuously, but triggered by specific events. An app switch. A click. A pause in typing. A scroll. When it can, the tool grabs the accessibility tree from your operating system; when it can't, it falls back to optical character recognition. Audio gets transcribed locally using Whisper Large-V3-Turbo, though you can route it to Deepgram's cloud service if you're willing to make that tradeoff.
All of it lands in a local SQLite database with full-text search. According to the project documentation, that works out to roughly 300 megabytes per eight-hour workday, compared to two gigabytes for continuous recording. Storage efficiency aside, the data stays put unless you explicitly flip on cloud features.
The app runs on macOS (both flavors—Apple Silicon and Intel), Windows 10 and 11, and Linux if you're willing to build from source. A REST API runs on localhost:3030, giving developers programmatic access to your timeline. There's also an MCP (Model Context Protocol) server that connects directly to Claude Desktop and Cursor, meaning those AI tools can query your work history without you copying and pasting context into a chat window.
It's event-driven rather than a fire hose, which seems to be where the efficiency gains come from.
Agents That Know What You've Actually Done
Screenpipe calls its automation layer "Pipes"—schedulable agents defined in markdown files. The repository includes samples: a meeting summarizer, a daily recap generator, something that breaks down how you spent your time, a prompt journal. These Pipes query the timeline database and execute actions on a schedule or when triggered.
The pitch emphasizes use cases that sound lifted from a productivity automation wishlist. Surfacing commitments buried in meetings and Slack threads. Identifying repetitive tasks that could become automated workflows. Maintaining a personal CRM based on who you've actually talked to rather than who you promised to follow up with.
The value proposition, stripped down, is that AI agents finally have access to ground truth—what you've been doing, not what you remember doing or claim you did.
Beaumont published a dedicated page on "computer-use agents" that frames Screenpipe as infrastructure. Accessibility trees, OCR results, transcripts, documented procedures—all in one queryable store for whatever agent-driven workflow tools other companies are building. It's a layer cake approach: Screenpipe handles memory, someone else handles the automations.
The Privacy Fine Print
Local-first comes with footnotes.
PostHog analytics and Sentry error tracking can be disabled in Settings → Privacy. Cloud transcription and sync features send data off your device only when you activate them. A security whitepaper—version 1.3, published earlier this year—spells out the architecture in detail.
Enterprise deployments get additional levers: per-pipe permissions defined in YAML, centralized configuration, SSO and SAML integration, audit logs, MDM support for Microsoft's Intune and SCCM. The system pauses recording when it detects DRM-protected browser content, a feature added in a recent update. You can blocklist specific apps, windows, or URLs.
Still, the inherent tension remains. One developer review noted the obvious: always-on capture raises privacy questions even when the data never leaves your laptop. Beaumont's response, reading between the lines, is that control over your own surveillance is better than handing it to someone else's servers.
The GitHub repository has generated substantial discussion, much of it debating exactly that tradeoff. Is local recording fundamentally safer, or just differently risky?
From Open to "Open Enough"
Screenpipe started life under an MIT license but switched to a source-available commercial license in early June. Personal use stays free; commercial use requires payment. The announcement referenced "hundreds of thousands of installs" and "nearly 20,000 stars" on GitHub—the star count is publicly verifiable, the install figures less so.
Current pricing runs $25 a month for the base subscription, $50 per seat for Pro, $150 per seat for Enterprise. Existing lifetime licenses were grandfathered. Beaumont reported $50,000 in monthly recurring revenue in a LinkedIn post earlier this year, up from $15,000—unverified figures that represent the founder's own marketing claims rather than independently audited financials.
The license change sparked the predictable debate about open-source purity versus commercial sustainability. Beaumont's position, inferring from the blog post, is that maintaining cross-platform desktop capture and building enterprise features costs money. Which is fair enough—though the shift from "open source" in early marketing materials to "source-available" in current documentation suggests some repositioning happened along the way.
A Gap in the Market, or Just a Niche?
Context matters here. Microsoft stumbled. Limitless got swallowed. There's daylight in the market for developer-friendly, locally-run capture tools, and Screenpipe moved quickly to occupy it—20,000 GitHub stars and a Y Combinator pedigree in hand.
The SDK opened recently, an Electron and JavaScript package that other developers can embed. The MCP server means Claude and Cursor users can already pipe their AI assistants directly into Screenpipe's memory. The changelog shows continuous updates: agent execution controls, selective sync, redaction features, UI performance improvements.
Whether enterprises will actually trust always-on recording, local storage or not, remains unanswered. But for developers building agent systems or teams documenting Byzantine workflows, Screenpipe offers something that didn't quite exist in this form before. The tool shipped its latest release one day before the Y Combinator launch.
Perhaps that timing says something about Beaumont's confidence. Or perhaps it just says he knows how to ship code.
Either way, the market will decide whether local-first surveillance is the privacy win Beaumont claims it is—or just a different kind of uncomfortable compromise.
