Sentient OS: An On-Device LLM That Reads Your Mac and Acts on It
An on-device LLM understands your entire life, then proactively offers to get your work done through computer use.
At a glance
- What is it?
- Sentient OS is an AGPL-3.0 macOS app that scans your local files, messages and notes each night with an on-device model, distills them into a markdown knowledge base, and offers computer-use actions the next morning. The design is unusually explicit about what stays local and what does not, but it also asks for deep access to your machine.
- Who is it for?
- Sentient OS is for Apple Silicon Mac users on macOS 15 or later who want a local-first personal knowledge base and are willing to grant an app broad access to their files, messages and notes. It is not for anyone on Intel Macs, on older macOS, or uncomfortable with a background process reading WhatsApp, iMessage and Apple Notes databases, even with the on-device classification step described in the README.
- Can I use it commercially?
- Yes, with strict conditions. AGPL-3.0 is a network copyleft licence: if people use a modified version over a network, for example as a hosted service, you must offer them its source code under the same licence.
- Is it still maintained?
- Yes. The repository last received commits 41 days ago.
- What is it written in?
- Mainly Swift, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The problem Sentient OS targets: assistants that forget you and wait to be asked
The README frames the problem in two parts. Existing AI tools know nothing about you, because, as it puts it, nobody can paste their entire life into a chat box. They also only help when summoned. Sentient OS is built for people who want an assistant that already has context and that volunteers work rather than waiting for a prompt. The target user is a single person on a single Mac, not a team, not a server deployment. The README states the app is free, open source and optimized for Apple silicon, with macOS 15 or later and 8 GB of RAM as the stated floor. The pitch rests on a compute argument: reading your entire life every day is expensive and privacy-sensitive in the cloud, but on your own chip it is, in the README's words, free, and nobody else's business. That is the whole justification for the local-first design, and it is the axis on which every other decision in the project turns.
The nightly pipeline: scan, classify, distill, then draft
The mechanism runs on a schedule. At 3 AM, according to the README, the app wakes the Mac (a closed lid is fine, and it falls back asleep afterwards) and reads what is new: files and screenshots, WhatsApp, iMessage and Apple Notes, decoded from the databases already on disk. Gmail and Google Calendar arrive through the user's own OpenAI connectors. Each item then passes through a local classifier, Gemma 4 E4B, which labels it keep, junk or sensitive. Junk is dropped. Sensitive is dropped with no summary, no log and no tombstone, which the README describes as zero trace. Only PII-stripped summaries of the keepers survive. Those summaries are then handed to a frontier model, reached through the user's own ChatGPT subscription, Open Router or LM Studio, which distills them into an Obsidian-style folder of plain markdown on the Mac. The README claims roughly 90 percent of the compute happens locally and the remainder goes to the frontier model. The output is a folder of notes the user can read, edit and delete individually, and the Knowledge window renders them in a Constellation View where notes are stars and wikilinks are threads.
Proactive cards and Sidekick: where the app stops being a note-taker
The knowledge base feeds two action surfaces. Overnight, a frontier model reads the night's findings and prepares a small set of morning cards: a drafted reply, a subscription cancellation caught before renewal, a piece of research. The user can read and edit a card, and one click fires it. The README is explicit that this click is the only thing that ever fires an action. The second surface is Sidekick, invoked by clicking the notch or by holding right Command to speak or tapping to type. It transcribes on-device, then performs computer use in the user's own apps and logged-in browser, with progress streaming in the notch. Because tasks are grounded in the knowledge base, the README argues Sidekick knows who people are and what promises were made to whom. That grounding is the difference the project claims between a generic agent and a proactive one. Whether the grounding holds up in practice is not something the README demonstrates; it is an architectural claim, not a measured result.
Getting it running: cask, Xcode project, and the model download
Installation is a Homebrew cask: brew install --cask sentient-os-labs/tap/sentient-os. Alternatively, download the DMG from the latest release. The README states the requirements plainly: Apple Silicon (M1 or newer), macOS 15 or later, 8 GB of RAM, and about 10 GB of free disk before starting, because the on-device model is a 3.7 GB download that needs room to land. Building from source means cloning the repository, opening Sentient OS macOS.xcodeproj in Xcode 26, and pressing Run; the on-device model downloads itself during onboarding. There is no account to create. The optional cloud MCP server is off by default. If enabled, the Mac seals the knowledge base with AES-256-GCM before anything leaves, and the README describes this as zero-access encryption, with the key living only on the Mac and in the user's private link, never on the project's servers. Turning sharing off deletes the cloud copy immediately, and a Mac that stops syncing causes the copy to self-destruct after 30 days. The relay is a separate open source repository, sentient-os-mcp.
Where the design gets uncomfortable: broad database access and a thin threat model
The app reads WhatsApp, iMessage and Apple Notes by decoding the databases already on disk. That is the same access pattern a forensic tool would use, and it is the price of the local-first promise. The README's answer is the Gemma 4 E4B classifier, which drops junk and sensitive items with no trace. That is a reasonable design, but it is a classifier, and classifiers have error rates the README does not quantify. A misclassified sensitive item is summarized and passed to the frontier model, which means it leaves the Mac. The README does not describe a review step before summaries reach the frontier model, and it does not describe what happens if the classifier is wrong. The second limitation is the frontier model dependency. The README says the last portion of compute taps a frontier model through the user's own ChatGPT subscription, Open Router or LM Studio. Without one of those, the distillation and the morning cards have nothing to run on. The local model reads and classifies; it does not do the final synthesis. So the app is free but not self-contained. A user who wants no external calls at all gets a scanner and a classifier, not the knowledge base the README leads with.
How it differs from Obsidian plus a local model, and from cloud assistant memory
The obvious comparison is Obsidian with a local LLM plugin. The output format is the same, plain markdown in a folder, and the README leans into that by calling the result Obsidian-style. The difference is the input side. Obsidian plugins act on notes you write or clip; Sentient OS writes the notes for you from files, screenshots, WhatsApp, iMessage, Apple Notes, Gmail and Google Calendar on a schedule. The second comparison is cloud assistant memory, the kind that accumulates facts about you inside a vendor's account. Sentient OS keeps the knowledge base as a folder on your Mac and, if you enable the MCP server, ships only ciphertext with keys that stay local. The trade is that you own the storage and the compute, and you also own the failure modes: a corrupt folder, a misconfigured connector, a model download that does not complete. The README does not describe a backup or export path beyond the fact that the notes are plain markdown, which is at least a format you can copy yourself.
Licence, maintenance and what the release cadence suggests
Sentient OS is AGPL-3.0. That matters if you plan to modify the code or run a modified version as a network service, because the licence's network clause applies to users interacting with it over a network. For an individual running the app on their own Mac, the practical effect is mostly about redistribution and modification, not daily use. The repository shows three releases in July 2026 (1.1 on the 20th, 1.2 on the 20th, 1.3 on the 25th) and a last push on 2026-08-06. That is a fast cadence, which cuts both ways: fixes arrive quickly, and so do changes to a system that reads your messages. The README does not describe a migration path for the knowledge base folder between versions, and it does not describe what happens to the 3.7 GB model download on upgrade. Anyone adopting this should treat the markdown folder as the durable artifact and the app as replaceable, because the folder is the part you can read without the app.
Who should adopt it, and what to check before you do
Adopt it if you are on an M1 or newer Mac with macOS 15 or later, you already pay for a ChatGPT subscription or run Open Router or LM Studio, and you want a personal knowledge base built from your own files and messages without sending raw data to a vendor. Skip it if you are on Intel, on an older macOS, or unwilling to let a background process decode your WhatsApp, iMessage and Apple Notes databases, even with the on-device classifier in front. Before installing, confirm free disk space against the stated 10 GB, decide which connectors you actually want, and leave the cloud MCP server off unless you need it. After the first nightly run, open the markdown folder and read what the classifier kept, because that folder is the only part of the system you can audit without the app. The README's closing claim is that every word of the pitch is verifiable in the codebase; that is the standard to hold it to, and the folder is where verification starts.
Editorial conclusion
Sentient OS is for Apple Silicon Mac users on macOS 15 or later who want a local-first personal knowledge base and are willing to grant an app broad access to their files, messages and notes. It is not for anyone on Intel Macs, on older macOS, or uncomfortable with a background process reading WhatsApp, iMessage and Apple Notes databases, even with the on-device classification step described in the README. Before adopting, verify the model download size and disk headroom on your machine, confirm which connectors you actually want enabled, and read the AGPL-3.0 terms if you plan to modify or redistribute the code.
Community notes