Overview
Sentient OS is a macOS on-device personal assistant from a two-person team in the Y Combinator Fall 2026 batch, released as open source under AGPL-3.0. The premise is specific: overnight, a local model reads the day's email, messages and files, and writes what it learns into plain Markdown files. What you find in the morning is not a summary but drafted replies, subscriptions about to renew and a report you promised to send.
Inference stays on the machine. The built-in model is Gemma 4 E4B running through the team's own LiteRT-LM fork with KV cache reuse, speculative decoding and custom k-quants, accelerated by Metal, which lets an Apple silicon Mac with 8GB of memory run the nightly pass. Your hardware handles roughly 90 percent of the compute, with the rest going to a ChatGPT or Claude subscription you already have, and the servers store nothing.
Three surfaces make up the product. Sidekick is summoned from the menu-bar notch and works across your apps and browser with its own cursor. Double Tap drafts a reply in whatever field you are typing in, following your context and phrasing, and never sends it. Overnight suggestions hunt for overdue replies, renewals and unexpensed receipts, preparing what can be prepared.
Key Features
- A knowledge base built at night: The local model reads mail, messages, files and meeting transcripts overnight and writes its understanding into a Markdown folder on your Mac. You can open, edit or delete any of it, and other AI tools can use it too.
- Sidekick takes over the task: Click the notch, say finish this for me, and Sidekick continues where you stopped using its own cursor across applications, while you move on to something else.
- Double Tap drafts replies: Double-tap the right Command key in any text field and it writes a draft in place, using the on-screen conversation and your writing style. Drafts wait for your approval.
- Nightly hunting for loose ends: It scans for unanswered messages, upcoming renewals and unfiled receipts, then prepares what it can so the morning is a review pass rather than a writing session.
- A custom on-device stack: KV cache reuse, speculative decoding and custom k-quants in a LiteRT-LM fork are what let an 8GB Mac run for hours at night. About 90 percent of the compute runs on the user's hardware.
- Swappable models, local data: Beyond the built-in Gemma 4 E4B, it can point at larger local vision models through LM Studio or compatible servers, and reach ChatGPT, Claude or OpenRouter endpoints when a task needs cloud capability. Connected local sources are analysed on-device.
Use Cases
- Anyone answering a dozen emails a day, using Double Tap to produce drafts that only need review.
- People with several half-finished tasks, handing the repetitive finish work to Sidekick.
- Keeping a portable personal knowledge base, since the Markdown folder moves with you and works with other tools.
- Users who need inference to stay local so sensitive material never leaves the machine.
- Forgetting subscriptions and bills less often, with a nightly pass surfacing what is about to renew.
- Small teams building something similar, since the AGPL source can be read and modified directly.
Pros
- Inference runs on the device, personal data stays local and the servers store nothing.
- Its understanding is written as Markdown, readable, editable and portable.
- An Apple silicon Mac with 8GB of memory is enough, a low hardware floor.
- Sidekick operates other apps with its own cursor, reaching the interface layer of computer use.
- The boundary stays explicit: drafts and actions wait for confirmation.
- AGPL-3.0 with commercial licensing available, so it works both as a tool and as a codebase.
Pricing
The app is open source under AGPL-3.0 and free for personal use, with commercial licensing negotiated separately. Running cost falls on local hardware: the team says the user's machine handles about 90 percent of the compute, with the remainder drawing on an existing ChatGPT or Claude subscription, so there is no separate cloud inference bill. Requirements are Apple silicon from M1, macOS 15.4 or newer, 8GB of memory and about 10GB of disk. Installation runs through a Homebrew cask or the GitHub release page.
Summary
Sentient OS bets on one judgement: having an AI understand your life overnight is unaffordable on cloud inference and awkward on privacy, but workable on the chip you already own. It makes that legible by writing the understanding into Markdown you can read, correct and carry elsewhere. A two-person team of teenagers with about a million dollars in pre-seed has over ten thousand users on it. For anyone who wants their data local, this is the most complete on-device path shipping today.
Version History
- Sentient OS launches with overnight on-device inference writing a Markdown knowledge base (2026-10-28): A YC Fall 2026 team open-sourced Sentient OS, a macOS on-device assistant under AGPL-3.0. It runs Gemma 4 E4B through a custom LiteRT-LM fork with KV cache reuse, speculative decoding and custom k-quants, so an 8GB Apple silicon Mac can work overnight while the user's hardware handles about 90 percent of compute. It reads mail, messages and files at night and writes understanding into an editable Markdown folder; during the day Sidekick operates apps with its own cursor and Double Tap drafts replies in place. Requires macOS 15.4 and M1 or newer, with over 10,000 users.