Louis Beaumont has a radical proposition: your computer should record everything. Every screen you glance at, every word you type, every offhand comment during a Zoom call. The idea sounds like a privacy nightmare—until you hear the rest. What if none of that data ever left your laptop?
That's the gamble behind Screenpipe, a Y Combinator-backed startup building what it describes as a "local-first AI memory system." The desktop application captures screen activity and audio continuously, storing it in a local SQLite database rather than shipping it to the cloud. Users can then search through their digital workday using natural language queries. Think Microsoft's controversial Recall feature, but cross-platform and designed for developers building AI agents that need real context about what you're actually doing.
Since its public debut on Hacker News last summer, Screenpipe's GitHub repository has accumulated 20.8k stars as of August 9, 2026. But the company's evolution from open-source project to commercial product reveals the messy realities of building privacy-focused infrastructure while trying to run a viable business.
Under the Hood
Screenpipe doesn't take constant screenshots—a common misconception. Instead, the system uses event-driven capture. Triggers fire when you switch applications, click, pause typing, scroll to a stop, or when the software detects visual changes on screen. Text gets extracted from accessibility trees first, with OCR as a fallback. Audio capture runs in parallel, transcribing conversations with Whisper locally or Deepgram if you opt for cloud processing.
The result is a searchable timeline of your workday. Natural-language queries like "what was that API endpoint Sarah mentioned last Tuesday?" surface relevant screen captures and audio transcripts. The company claims the application consumes between 5 and 10 percent of CPU resources, 0.5 to 3 GB of RAM, and roughly 20 GB of storage monthly.
For developers, Screenpipe exposes a localhost REST API and an MCP server that integrates with Claude, Cursor, and ChatGPT. The real utility, according to the company, comes from "Pipes"—scheduled agents written in markdown that can generate end-of-day recaps, standup summaries, time breakdowns, or meeting notes. Think cron jobs for your workday, except they're powered by everything the system has seen and heard.
The software runs on macOS (both Apple Silicon and Intel), Windows 10 and 11, and Linux via source builds. As of this summer, the company was shipping multiple releases in a single week, addressing database lifecycle issues, Windows overlay bugs, and enterprise scheduling problems.
The Security Question
If you're going to record everything, security architecture matters. Screenpipe published a security whitepaper this past spring detailing its approach: local-first modes by default, token authentication for the localhost API, CORS restrictions, AES-256-GCM for secret storage, and ChaCha20 encryption for data at rest in what the company calls the "Vault."
The most recent addition is a 278 MB ONNX model for on-device personally identifiable information redaction. Announced in late May, the INT8 quantized model was benchmarked against a dataset the company dubbed "ScreenLeak." Screenpipe now positions redaction as enabled by default, though the documentation shows some inconsistency about whether the local API requires authentication out of the box—a detail that remains unclear across their published materials.
For enterprise deployments, Screenpipe offers admin controls, team encryption, central configuration, and mobile device management policy enforcement. The architecture breaks down into 13 Rust crates and integrates with Notion, Slack, Google Calendar, Obsidian, Toggl, Salesforce, HubSpot, Zoom, PostHog, and Sentry.
Still, even local-first software presents risks. A Check Point security assessment from earlier this year categorized Screenpipe's capture of sensitive on-screen and audio data as a potential risk—not unique to this tool, but worth noting for IT teams evaluating it.
The License Shift
Screenpipe started as an MIT-licensed open-source project. That changed in early June when the company announced a shift to the "Screenpipe Commercial License"—source-available, free for personal and non-commercial use, but requiring a paid license for commercial deployments.
The decision drew swift pushback. Reddit moderators removed posts claiming Screenpipe as "free software," and Hacker News threads—some dating back to 2024—had already criticized the company's promotional tactics and ambiguous positioning around open source. Some earlier versions remain under MIT, but the current codebase is not open source by Open Source Initiative standards.
Pricing has shifted just as dramatically. The company moved to a subscription model, with prices confirmed as of August 9, 2026. In late July, a free tier arrived alongside Gmail integration for all plans. The GitHub README now lists subscriptions "starting at $25/month," though the public pricing page redirects to onboarding, suggesting the model is still in flux. Existing lifetime licenses remain valid.
Carving Out a Market
Screenpipe positions itself against two camps: Microsoft Recall (Windows-only, screen-only, plagued by security concerns) and cloud-based meeting tools like Granola and Otter.ai (limited to specific contexts, data sent to third parties). The company maintains a comparison hub on its website, systematically differentiating itself from Limitless (formerly Rewind), Pieces, Omi, Scribe, and others.
The pitch for developers is clearer. Screenpipe wants to be infrastructure. If AI agents are going to work on your behalf, they need context—not just what's in your calendar or email, but what you've actually been doing, seeing, and saying. The MCP integration lets tools like Claude access that context without shipping your data to Anthropic's servers.
Third-party projects have taken notice. Community lists and GitHub discussions position the tool as a privacy-conscious option in the emerging "recall" category.
Funding, Sort Of
Screenpipe participated in Y Combinator's Summer 2026 batch. Beyond that, funding details remain murky. A LinkedIn post from investor Choongsoon Bae mentions a $2.8 million pre-seed round at a $40 million valuation, listing Founders Inc., Seed Club Ventures, and LG Technology Ventures among backers—though this remains unverified by primary sources. VCBacked, a startup aggregator, echoes the $2.8 million figure. But no official press release has surfaced, and the company's "Invest in Screenpipe" landing page gates details behind a form.
Older blog posts from early 2025 reference "raising $5M" and weekly active user updates, but those predate the license change and current product architecture. Context, perhaps, but not current fact.
What Comes Next
The product is shipping at speed. Recent updates include battery-conscious power tiers, mobile selective sync, reliability hardening to reduce capture gaps, editable speaker labels, and the free tier rollout. The company's blog—essentially a chronological changelog—shows the pace: YC announcement, power tiers and ONNX redactor, license change, reliability and mobility updates, free tier and Gmail. It's the cadence of a team racing toward product-market fit.
Whether enterprises will embrace the pitch is another question entirely. Recording everything employees do is a hard sell even when the data stays local. The tool works best for individuals who want a searchable memory layer and developers building agent workflows. Teams and IT departments face thorniest compliance, consent, and cultural concerns—not all of which can be solved with encryption.
Screenpipe has built what it set out to build: a technically impressive system for capturing and querying your digital workday, with serious thought given to privacy architecture. Whether that's enough to overcome the inherent discomfort of constant surveillance—even self-surveillance—remains unsettled. The company is betting that in a world of cloud-based AI assistants, local-first wins on principle. The GitHub stars suggest developers agree. The silence around enterprise adoption? That suggests everyone else is still deciding.
