A new wave of "always-on" personal agents wants to read your whole digital life while you sleep. Sentient OS, an open-source Mac app from a Y Combinator F26 team, is the latest, and it makes an unusually specific promise: a local model analyzes your mail, messages and files overnight and writes what it learns into Markdown files you own. The launch is making the rounds on X today, with developers highlighting local models as a cheap way to run always-on proactive agents.
The pitch is appealing, but "on-device" hides a lot of detail. This post goes through the project's GitHub README and release notes to separate what runs locally from what does not, and shows how to try it safely. One note on timing: the repository already has a 1.3 release that third-party trackers list as about three months old, so today's buzz is a launch push for an existing open-source project, not a brand-new codebase.
TL;DR: Sentient OS in one table
| Question | Answer (per the README) |
|---|---|
| What is it? | Open-source personal AI for Mac that builds a knowledge base and offers to do work via computer use |
| Who makes it? | Sentient OS Labs. The README says "Backed by Y Combinator"; the F26 batch and two-person San Francisco team come from press reports (DoNews) |
| Built-in model? | Gemma 4 E4B via a customized LiteRT-LM integration with Metal acceleration |
| Other models? | Larger local vision models via LM Studio or compatible servers; hosted models via ChatGPT or Claude account, OpenRouter or any compatible endpoint |
| Requirements | Apple silicon (M1+), macOS 15.4+, 8 GB RAM, about 10 GB disk |
| License | AGPL-3.0, commercial licensing available separately |
| Install | brew install --cask sentient-os-labs/tap/sentient-os or GitHub release |
| Traction | 644 stars and 61 forks when we checked the repo |
What does Sentient OS actually do?
The README describes three features:
- Sidekick: invoked from the menu-bar notch, it works in your apps with its own cursor, clicking, typing and attaching files.
- Double Tap: drafts a reply in the field you are typing in, using the on-screen conversation and examples of your writing. Drafts are never sent automatically.
- Proactive suggestions: overnight, it hunts for loose ends such as overdue replies, upcoming renewals and expense receipts, and prepares drafts for your review in the morning.
Underneath is a knowledge base: a plain folder of Markdown files you can read and edit. That choice matters. It makes the agent's memory inspectable and portable, the same philosophy behind Obsidian-style vaults that other agent tools are adopting for codebases.
How "local" is it really?
This is the key question, and the answer has layers.
Local Mac agent Sentient OS runs its built-in model on an Apple silicon laptop
Local by default: selected sources, including files, Messages, WhatsApp, Notes, Mail and Calendar, are first analyzed on your Mac. The built-in model is Gemma 4 E4B running through a customized LiteRT-LM integration with Metal acceleration, which is how it fits an 8 GB machine.
Not necessarily local:
- After the first pass, "useful summaries then go to your chosen AI, local or hosted," which organizes them into the knowledge base. If you choose ChatGPT or Claude, those summaries leave the machine.
- Double Tap's default route passes screen and knowledge context through Sentient's relay to OpenAI under Zero Data Retention. You can switch to a local provider.
- Browsing and sending email need the internet, even with local AI.
- Optional cloud MCP sharing encrypts the knowledge with AES-256-GCM before upload.
Third-party descriptions disagree on how much compute stays on the device. One aggregator claims the user's hardware does about 90% of the work, while a Chinese-language report from DoNews says the app leans on users' existing ChatGPT and Claude accounts for large-model inference. The README's wording, with a built-in local model plus a choice of provider, reconciles the two: the split depends on your settings. We could not verify the 90% figure from a primary source, so treat it as unconfirmed.
What changed in release 1.3?
The 1.3 release notes say the app no longer requires a ChatGPT subscription. It can point at any endpoint that speaks the OpenAI Responses API, with built-in support for OpenRouter, LM Studio and self-hosted models. Your device can run the private knowledge base, morning cards, Sidekick and computer use without an OpenAI account. The remaining catch: the Gmail and Google Calendar connectors work only through a ChatGPT account connector for now. The release also made computer-use setup work again on fresh installs and made it faster and Codex installation more reliable.
How does it compare with other local Mac agents?
Local-first agents are crowding in. Perplexity's fully local Mac mode with Qwen 3.6 needs far more memory, 36 GB in the post's title. Underdog's private personal AI takes an on-device privacy architecture approach, and Google AI Edge's Foresight does local meeting notes. Sentient OS differs by being proactive: it runs on a schedule, not on prompt, and it acts through computer use.
That second point is also the risk. Meta Muse's access to Messages already raised Full Disk Access questions, and a desktop workspace like OpenBot's danger-full-access mode shows how licensing and permissions vary across tools.
What are the risks and limits?
The README is honest about several limits:
- macOS only. Windows is not available yet.
- Plugged-in Mac required. Overnight processing needs the app running in the menu bar, and the first analysis can take hours.
- Full Disk Access. The app needs it to read the local databases you enable.
- No per-click confirmation. Sidekick does not ask before each click, so you are authorizing it to complete the task.
- Feature dependencies. Some features depend on your connected account or plan.
- AGPL-3.0. Great for transparency, but companies embedding it in network services have obligations; commercial licensing exists.
Beyond the README, remember the project is young: a small team, a few hundred stars and, at the time we looked, an empty Releases section on the main page capture despite the 1.3 notes existing on a mirror. We did not find an independent security audit. "Open source" lets you inspect it, but it does not by itself make it safe.
How to try it more safely
This is our suggested approach, not the vendor's instructions:
- Start on a secondary profile or machine if you can, with only one source enabled, such as Notes.
- Choose a local provider (LM Studio or the built-in Gemma 4 E4B) for both the knowledge base and Double Tap if privacy is the goal, and confirm in settings that the OpenAI relay is not in use.
- Read the Markdown output after the first night. It is the clearest view of what the agent inferred about you.
- Never leave unattended agents with payment or send permissions until you trust the drafts.
- Watch network traffic with a firewall tool for the first few days.
If you run other agents that act on your machine, AgentBeam is the agent security platform from the explainx.ai team that stops AI agents before they take dangerous actions.
What are people asking about it?
Is it free? The app is open source under AGPL-3.0, and the built-in model needs no account, so the local path has no per-token cost. If you choose ChatGPT, Claude or OpenRouter, you pay whatever that provider charges. The README mentions separate commercial licensing but we found no public price list.
Is it the same as the other "Sentient"? No. Search results mix it up with Sentient AI, the open-source AGI company that raised $85 million. They are unrelated.
Who is behind it? Third-party reports name Aditya Vellanki as CTO and Jesai Tarun as CEO of a two-person San Francisco team that reportedly raised over $1 million pre-seed. We could not confirm those figures from the company itself, so read them as reported, not verified.
How big is the community? The repository showed 644 stars, 61 forks and 5 watchers when we looked. An aggregator claims 2,000+ users within 48 hours of a first Reddit post, but that site looks low-quality and we could not confirm the number.
Which setup fits which person?
| If you want | Pick this setting |
|---|---|
| Maximum privacy | Built-in Gemma 4 E4B or a larger LM Studio model, local provider for Double Tap, a single source enabled |
| Better drafts, less privacy | A ChatGPT or Claude account as the organizing model, accepting that summaries leave the Mac |
| Gmail and Calendar | The ChatGPT account connector, the only route in 1.3 |
| Lowest cost | Local models only; the trade-off is slower overnight runs on 8 GB machines |
Why this matters
Small open models such as Gemma 4 E4B are now good enough to do the bulk reading work on a base-model laptop, which is what makes overnight, unlimited inference economically plausible: there is no per-token bill. Our look at how a tiny fine-tuned model beat frontier models on a task points the same way. The open question for Sentient OS is not capability but trust: whether a two-person team can earn permission to read everything on your Mac.
Related reading
- Perplexity Computer, fully local on Mac
- Underdog private personal AI
- Full Disk Access and AI agents
- Google AI Edge Foresight for local meeting notes
- OpenBot desktop workspace
- Graphify: codebase knowledge graph skill
Details are accurate as of October 27, 2026, from the project's public README and release notes. Check the repository for changes.