# Sentient OS: an on-device LLM that reads your Mac every night and drafts your work

> Sentient OS is a free, AGPL-3.0 macOS app for Apple silicon that runs a local model over your files and messages at 3 AM, writes a markdown knowledge base, and offers one-click drafts in the morning. The trade-offs are in the hardware requirements and the frontier-model dependency.

**Sentient-OS-Labs/sentient-os** — An on-device LLM understands your entire life, then proactively offers to get your work done through computer use.

- Repository: https://github.com/Sentient-OS-Labs/sentient-os
- Website: https://sentient-os.ai
- Stars: 526 · Forks: 44
- Language: Swift
- License: AGPL-3.0
- Published: 2026-09-10 · Updated: 2026-09-10 · Language: en
- Canonical page: https://hysenlabs.com/projects/sentient-os-labs-sentient-os

## The problem Sentient OS picks: assistants that know nothing and wait to be asked

The README states the case plainly: every AI you use today knows nothing about you, and it only helps when summoned. Those are two separate failures. The first is a context problem. Nobody can paste their entire life into a chat box, so the model answers from general knowledge and guesses at the specifics. The second is an initiative problem. Even a model with perfect context sits idle until you type something.

Sentient OS attacks both at once, and the README is explicit that the cost is what forces the design. Reading your entire life every day is expensive inference in the cloud, so the project moves the bulk of it onto the Mac. The intended user is a single person on one machine, not a team: there are no accounts, and the README says raw personal data never leaves the Mac. If you want a shared knowledge base for a group, this is the wrong shape of tool.

## The 3 AM pipeline: bouncer, summaries, then a frontier model

The architecture splits compute roughly 90/10 between your Mac and a frontier model you already pay for. The README describes the local pass in detail. At 3 AM the app wakes the Mac (a closed lid is fine, and the machine falls back asleep afterward) 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 your own OpenAI connectors.

Every item then passes what the README calls an on-device bouncer. Gemma 4 E4B runs locally and rules each item keep, junk, or sensitive. Junk is dropped. Sensitive is dropped harder, with no summary, no log, and no tombstone. Only PII-stripped summaries of keepers survive that stage.

The finale is the part worth scrutinising. Those summaries go to a frontier model, which distills them into an Obsidian-style folder of plain markdown on your Mac. Local inference does the reading; the remote model does the synthesis. That is a deliberate trade, and it means the quality of your knowledge base depends on the subscription you connect.

## Installing Sentient OS on macOS and seeing the first knowledge note

The README gives two install paths. The fastest is Homebrew:

```bash
brew install --cask sentient-os-labs/tap/sentient-os
```

That pulls the cask from the project's own tap. The alternative is downloading the DMG from the latest release page. Requirements are stated up front: an Apple Silicon Mac (M1 or newer) on macOS 15 or later, 8 GB of RAM, and about 10 GB of free disk before you start, because the on-device model is a 3.7 GB download and needs room to land.

If you prefer to build it yourself, the README says to clone the repository, open `Sentient OS macOS.xcodeproj` in Xcode 26, and press Run. The on-device model downloads itself during onboarding either way, so the first launch is not instant.

Once onboarding finishes, the thing to look at is the knowledge base folder. It is plain markdown, and the README stresses that you can read, edit, and delete it note by note. The Knowledge window also has a Constellation View, where notes appear as stars and wikilinks as threads. The first real check is whether the folder contains notes you recognise from your own files, because that is the whole product working.

## Sidekick, the notch, and computer use in your own apps

Sidekick is the interactive half. Anywhere on the Mac you click the notch, or hold right Command to speak and tap it to type. The notch opens, transcribes you on-device, and then computer use takes over, operating your own apps and your own logged-in browser with progress streaming in the notch. The README's examples are "finish this for me", "reply with the update from Sarah", and "put these ingredients in my cart".

The grounding claim is the interesting one. Because tasks run against the knowledge base, Sidekick is meant to know who the people in your life are, what "the usual" means, and what you promised whom. That is a real dependency, not a feature bullet: if the nightly pass produced thin notes, the agent has thin context to work from. The README positions the difference as one between an agent and a proactive agent that knows you, and the knowledge base is what carries that difference.

## The cloud MCP server, and why it is off by default

The knowledge base can be shared with ChatGPT and Claude, including phone apps, over a cloud MCP server. The README is careful about the default: it is optional and off by default. When you turn it on, your Mac seals the knowledge base with AES-256-GCM before anything leaves, and the project describes this as zero-access encryption, with the key living only on your Mac and in your private link. The relay holds ciphertext and no key. The relay is a separate open source repository, sentient-os-mcp.

The lifecycle rules matter more than the cryptography claim. There is no account, so turning sharing off deletes the cloud copy on the spot, and if your Mac simply stops syncing, the copy self-destructs after 30 days. The comparison the README draws is with a normal cloud backup, where the company keeps your keys and can be compelled to decrypt. That is the argument, and it is a design argument rather than a proof. The MCP path is the one place where data leaves the machine, so if your threat model does not include a relay holding ciphertext, leave the toggle alone.

## What Sentient OS will not do for you

The requirements are the first hard boundary. Apple Silicon only, M1 or newer, macOS 15 or later. There is no Intel build and no Linux or Windows story in the README. The 8 GB RAM claim comes with a caveat the project states itself: the on-device model was tuned until it fit, which tells you the memory budget is tight rather than generous.

The second boundary is the frontier model. The README names three routes, your ChatGPT subscription, Open Router, or LM Studio. The local pass is free and private, but the distillation step is not self-contained. A user with no frontier subscription gets the reading half without the synthesis half, and the README does not describe a purely local fallback for that stage. This is not a project you can run fully offline and get the advertised result.

The third is data access. Reading iMessage, WhatsApp and Apple Notes means the app touches the databases those applications keep on disk. Sensitive items are dropped with no tombstone, but the bouncer's judgement is a model's judgement, and the README does not describe an audit log of what the bouncer classified as sensitive. If you need to prove what was excluded, this design does not give you that.

## How it compares with an Obsidian vault plus a local model

The nearest alternative is assembling it yourself: a local model through Ollama or LM Studio, a script that watches a folder, and an Obsidian vault as the store. That route gives you full control over every stage, including a fully local synthesis step, and it is the honest comparison because the output format is the same plain markdown.

The difference in approach is where the intelligence sits. A DIY pipeline usually means you decide what to ingest and when, and you write the prompts that summarise it. Sentient OS inverts that: it decides what is new, filters it through a local bouncer, and hands the distillation to a frontier model on a nightly schedule. You get less control and more automation, plus the proactive cards and Sidekick's computer use, which a folder-watching script does not attempt. You also inherit a dependency on a paid frontier model and on the project's own pipeline staying correct. If you enjoy owning the pipeline, the DIY route is better. If you want the pipeline to exist without you building it, that is the gap Sentient OS fills.

## Licence, maintenance and the cost of upgrading

The repository is licensed AGPL-3.0, with a LICENSING.md file alongside it, and the README calls the app free forever. AGPL-3.0 is a strong copyleft licence, and the practical implication for anyone modifying and distributing the app, or running a modified version as a network service, is that source obligations attach. The repository also carries a CLA.md and a CONTRIBUTING.md, so contributions go through a contributor licence agreement. This is a description of the files present, not legal advice; read LICENSING.md before you build on the code.

The last push to the repository was on 2026-08-06, and the repository is not archived. Three releases are listed for July 2026: 1.1 on 2026-07-20, 1.2 later the same day, and 1.3 on 2026-07-25. That cadence suggests active work, but the README does not document an upgrade path, a migration story for the knowledge base format between versions, or a rollback procedure. The knowledge base is plain markdown, which limits the blast radius of a bad release, but the app state around it is undocumented. Budget for re-running onboarding after a major version if the on-device model changes.

## Conclusion

Sentient OS is for Mac users on Apple silicon who already pay for a frontier model and want a local knowledge base built from their own files without an account. It is not for Intel Macs, Linux, Windows, or anyone unwilling to grant an app access to iMessage, WhatsApp and Notes databases. Before adopting it, verify three things: that your Mac is M1 or newer on macOS 15, that you have roughly 10 GB free for the 3.7 GB model download, and which frontier provider you will connect, since the nightly distillation step routes PII-stripped summaries through your own ChatGPT subscription, Open Router, or LM Studio.

## FAQ

### What are the system requirements for Sentient OS?

An Apple Silicon Mac (M1 or newer) running macOS 15 or later, with 8 GB of RAM and roughly 10 GB of free disk space. The on-device model is a 3.7 GB download that lands during onboarding.

### Does Sentient OS send my data to the cloud?

The README says raw personal data never leaves your Mac, and the local bouncer drops sensitive items with no summary or log. The exception is the optional cloud MCP server, off by default, which syncs an AES-256-GCM sealed copy of the knowledge base to a relay that holds only ciphertext.

### How do I install Sentient OS on a Mac?

Use the Homebrew cask from the project's tap, or download the DMG from the latest release page. Building from source means cloning the repository, opening Sentient OS macOS.xcodeproj in Xcode 26, and pressing Run.

### Where is the Sentient OS knowledge base stored?

It is an Obsidian-style folder of plain markdown on your Mac, which the README says you can read, edit, and delete note by note. The Knowledge window also renders it as a Constellation View with notes as stars and wikilinks as threads.

## Sources

- [License: AGPL-3.0](https://github.com/Sentient-OS-Labs/sentient-os/blob/main/LICENSE)
- [Project website](https://sentient-os.ai)
- [README](https://github.com/Sentient-OS-Labs/sentient-os/blob/main/README.md)
- [Releases](https://github.com/Sentient-OS-Labs/sentient-os/releases)
- [Sentient-OS-Labs/sentient-os on GitHub](https://github.com/Sentient-OS-Labs/sentient-os)

---

Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/sentient-os-labs-sentient-os