Rewisp

Rewisp is a free, open-source macOS menu bar app built by Yashmit Bhaverisetti, a student at UCSD. It quietly captures text from your screen as you work — never the pixels, only what was written — and lets you ask questions about anything you've seen, answered by on-device AI. The project launched on Product Hunt on 20 July 2026, finishing #5 product of the day with 187 upvotes according to the project's own README.

What It Is

Rewisp sits in the category of ambient screen memory tools — a niche pioneered by products like Rewind and Microsoft Recall, but built with a privacy-first, fully local, open-source approach. Every time you switch apps, open a page, or settle after a scroll, Rewisp takes a "wisp": it OCRs the screen in memory using Apple's Vision framework, stores only the recognized text in a local SQLite database, and immediately discards the screenshot. You summon the panel with ⌘⇧Space and ask in plain English. The app requires macOS 15+ and Apple Silicon.

Privacy Architecture

The project's core design constraint is that screenshots never touch disk. Key privacy properties, as stated on the project homepage and README:

• OCR happens in memory; the image is discarded before anything is saved
• All data stays in a single SQLite file in ~/Rewisp, encrypted at rest with SQLCipher (AES-256)
• Answers can run entirely on Apple's on-device model — nothing leaves the Mac unless you ask a question that escalates to Claude, Gemini, or ChatGPT
• A kill list fully pauses capture for messaging apps, password managers, banking sites, and private browser windows
• Card numbers and SSNs are stripped via Luhn/SSA validation before storage
• A "Forget" button deletes the last 10 minutes; all data expires after 6 months
• The Touch ID Vault stores personal details (resume, addresses) behind a fingerprint and refuses credentials at ingest

Reasoning Features Beyond Search

Rewisp goes beyond simple retrieval. The README describes several reasoning-layer capabilities:

• Promise tracking: catches commitments like "I'll send it Friday" as they scroll past, reads deadlines like "by EOD" or "next week", and surfaces a reminder on the due morning
• Page diffing: keeps every version of a page as text and can show exactly what was added, changed, or removed since you last looked
• Semantic search: local embeddings fused with keyword search find pages by meaning, not just shared words; a trigram signal handles OCR typos
• Numbers over time: a label and number seen repeatedly (weight, grade, price) becomes a tracked sparkline
• Forgetting model: learns how you personally forget, strengthens facts you look up repeatedly, and lets the rest fade
• MCP integration: exposes your memory to Claude Desktop, Cursor, VS Code, or any MCP client as read-only local tools via stdio — no network listener, Vault excluded

Smart Answer Routing

Rather than requiring an API key, Rewisp chains through models you already have access to. The order, as described on the homepage: Apple on-device model tries first; if the answer is thin, it escalates to Gemini (free tier), then Claude, then ChatGPT — using subscription access, not billed API keys. A nightly digest at 9 PM summarizes the day, flags unfinished items, and proposes facts to remember — each requiring explicit user approval before being kept.

Update: v0.27.0 — Privacy + Smarter Memory

The latest release is v0.27.0, published 22 July 2026, titled "Rewisp 0.27.0 — privacy + smarter memory." The repository was created on 9 July 2026 and has seen active development, with the last push on 22 July 2026. The project notes that it is not yet signed with an Apple Developer ID certificate, which means macOS will ask users to re-enable screen recording after each update — a known limitation the README addresses directly. The project is actively seeking contributions; the README notes the first outside PR fixed two shipped bugs.

Open-Source Deployment Model

Rewisp is MIT-licensed and distributed as a .dmg via GitHub Releases. It ships its own Python runtime and sets up its own background helper on first launch — no Terminal commands or Python installation required. Updates are one-click from inside the app. The codebase is split between a Python daemon (rewisp/) handling capture, OCR, storage, and retrieval, and a SwiftUI menu bar app (ui/Sources/) for the interface. A pytest suite of 223 tests is included. The project is built and maintained by one person and is explicitly seeking grants and collaborators.