Last updated: 2026-08-18
Local Mind is designed so that your prompts, conversations, and files never leave your Mac. All AI inference runs locally on your device using Apple's GPU. We operate no server that receives your prompts, conversations, or documents, and the app works fully offline once a model is on disk.
One clarification, because the app can now do something that makes "your Mac" ambiguous: if you choose to let other people use your model (see If you serve others below), their prompts travel to your Mac and are processed there. Nothing about that sends anything to us — but you become the operator of that service, and this policy cannot speak for the relationship between you and the people you serve.
Nothing, unless you switch something on. Local Mind has no user accounts, no advertising, and no background analytics. Your prompts, conversations, responses, and files are never collected and never leave your Mac — the inference engine that sees them is network-locked and cannot reach the internet at all.
Two kinds of data can be shared, and only by an explicit opt-in: anonymous performance measurements, and specific preferences you choose to submit. Both are described below, both start off, and both show you the exact data before any of it is sent.
Optional: share anonymous performance data (upload, OFF by default — direct-download build only). In the version distributed outside the Mac App Store, you can choose to turn on sharing so the app contributes how fast models run on your Mac. What it sends is content-free by construction: a model identifier, coarse hardware (chip family, RAM tier, GPU core count, macOS major version), and measured throughput (tokens/second). It never includes your prompts, responses, files, any account or device identifier, or your Mac's exact board model. Before anything is sent you can preview the exact data in the app, and you can turn it off at any time. This stays off unless you explicitly enable it. The Mac App Store version has no upload capability at all — it never sends performance data.
Optional: preferences you choose to submit (per item, reviewed before sending — direct-download build only). You can offer back a few of your own settings, so the app can support what people actually use: ranking sites you added, the names of custom agents you configured, the command line of those agents, and the paths of local model files you run. This is deliberately not one on/off switch:
Your prompts, conversations, files and run history are never in this list. The Mac App Store version has no upload capability at all, so it cannot submit any of this.
Optional: community reference speeds (download, ON by default). The app can download the aggregated result of that shared data — anonymous median throughput per model and coarse hardware class — so its on-device speed estimates are grounded in real measurements instead of guesses. This is a read-only fetch that sends nothing about you. It can be turned off in the app; turning it off simply means the app shows its own estimates only.
https addresses are accepted. From that point it is an ordinary visit made by your browser under its own rules, so the site sees whatever your browser normally reveals (including your IP address); the app sends it nothing about you and is not involved in the session.https://pu-rs.org/v1/news.json at most twice an hour and cached on disk. The list is curated by us from a fixed set of sources — nothing user-authored — and it carries links only, never prices. The request carries no information about you beyond what any HTTPS fetch discloses (your IP address and a cache validator), and the panel can be switched off next to the sharing controls.The app functions fully offline once a model is downloaded.
These are not just promises — they are enforced by the operating system and by how our backend is built, and can be checked independently:
com.apple.security.network.client). macOS therefore refuses it any outbound internet connection at all, so your content cannot leave the device through it. It holds only the permission to accept incoming connections (com.apple.security.network.server), which is what lets it answer the on-device API. This is verifiable on the installed app: codesign -d --entitlements - "Local Mind.app/Contents/MacOS/ds4-server" shows the incoming-connection permission and no network-client permission.Beyond the signature, the engine locks itself down at runtime: once it has opened its listening socket it irreversibly drops its own ability to make outbound connections, then verifies the lock by attempting one and confirming the system refused it. The app reports what the engine says about itself rather than asserting it — the engine only claims to be locked after that attempt fails.
The app can issue API keys so that people you choose — colleagues, customers, friends — can use the model running on your Mac, whether you charge them or not. This section is about what that means, because it is the one feature that changes who the data belongs to.
Their prompts come to your Mac. When someone uses a key you issued, their prompt travels over the network to your machine, is processed there, and the answer is sent back. That is the whole point of the feature, and it is worth stating plainly: you are handling other people's data.
We are not part of it. We receive nothing from this — not their prompts, not their responses, not their identities, not the fact that it happened. There is no account, no telemetry and no reporting attached to it. The engine that reads their prompts cannot reach the internet at all, and the keys you issue are stored only as fingerprints on your Mac: not by us, and not recoverable even by you.
You are the operator, and that carries obligations we cannot discharge for you. Depending on who your users are and where they live, you may have duties to tell them what you do with their data, to keep it only as long as needed, to answer their requests about it, and to have a lawful basis for processing it. Those duties are yours, not ours — this policy governs what we do, and we are not a party to your relationship with the people you serve. If you are serving anyone beyond yourself, especially in a jurisdiction with data-protection law, take your own advice on it.
What we do to keep it from happening by accident. Serving beyond your own Mac is off by default and resets to off at every launch. Exposing the service to the wider internet refuses to start unless key-checking is on, so there is no configuration in which the app knowingly serves strangers with no key at all. Keys can be given an expiry and can be revoked at any time, which cuts off the holder immediately.
Local Mind does not see, log, or moderate what your users send. Neither do we. If you need a record of that, or need to review it, that is something you would have to build.
The app only accesses files you explicitly choose (via the system open panel) or the model files it downloads into its own container. The Mac App Store version runs inside Apple's App Sandbox.
We do not sell your data, and nothing you choose to share with us is passed on to anyone else. Two ordinary third-party interactions are worth naming, because they are visits you initiate: downloading a model contacts the publisher that hosts it (e.g. Hugging Face), and clicking a leaderboard link opens that site in your own browser. Each sees what any HTTPS request reveals, such as your IP address; neither is told anything about you by the app. Model weights you download are governed by their own publishers' terms and licenses.
The app is not directed at children. It collects nothing from anyone by default, and the optional shares described above require a deliberate opt-in.
If this policy changes, we will update this page and the "Last updated" date.
Questions: yijun.yu@gmail.com