Features Models Benchmark Ranking FAQ

Legal

Privacy Policy

Last updated: November 9, 2026 — 2.3.1

MLXHub runs language, voice and image models on your device, with the user-requested network exceptions described below. Analytics and diagnostics use a persistent random identifier and include model, performance and IP-derived approximate location data. They can be turned off in Settings. Publishing benchmark results is separate and optional: it makes device measurements public through CloudKit, linked to your iCloud account.

What the app accesses

Model files

MLXHub downloads model weight files and configuration files from Hugging Face (huggingface.co) when you tap “Download” on a model. These files are stored in the app's private storage, excluded from iCloud backup, and are never uploaded anywhere.

Alternate model source

In Settings, you can switch model downloads to ModelScope, a Hugging Face alternative hosted in China that helps where Hugging Face doesn't resolve. If your device's region is set to mainland China and you haven't chosen a source yet, MLXHub starts with ModelScope; everywhere else it starts with Hugging Face. You can change the source at any time. While ModelScope is the source, the repository name or search term you type, plus the standard download request for the model files, go to modelscope.cn — no personal information and no app identifier. The rest of the app keeps using Hugging Face.

Local network

When you enable the LAN server (opt-in toggle in the Server tab), the app opens a TCP listener on your local network. This listener is:

  • Reachable only from devices on the same Wi-Fi network.
  • Not exposed to the internet (no port-forwarding, no UPnP, no relay).
  • Advertised via Bonjour on the local network only.
  • Optionally protected by an API key you generate in the app.

The LAN server does not send request or response content to analytics. The Server screen can show client IP addresses and request counts, cleared when it stops. Use of server features and successful first inference can generate the limited analytics described below; no client IP or request text is included in those app events.

Embeddings, reranking and Ollama-compatible endpoints

The LAN server also answers embedding and reranking requests, and exposes your models through endpoints compatible with the Ollama API so that existing tools can talk to it. These endpoints use the API-key protection configured for the server. The text a client sends is turned into vectors or scores on your device, by a model you downloaded — there is no remote provider — and the assistant's memories (see below) are never visible from any of them.

Distributed inference

MLXHub can split a single model across several of your own devices on the same local network. Devices are paired explicitly, with a short code you confirm on both, and the connection between them is encrypted. Because the model runs jointly, the prompt for each turn — including the system prompt, your personality text, any attached document and the assistant's memory block — travels to the devices taking part, and a device that routes its chat through another one sends it the conversation text inside that encrypted connection. This data goes only to devices you paired, never leaves your local network, and there is no intermediary server or relay. The devices also exchange performance figures (tokens per second, network latency) with each other, not with us.

Attachments

You can attach documents, images and videos to a message. Text is extracted from documents on your device; videos are copied into the app's private storage, excluded from iCloud backup, and sampled into frames locally — no audio is captured or processed at any point. Photos and videos are chosen with the system picker, which does not require access to your whole library, and you can take a photo with the camera to attach it. Attachments are not sent to analytics. Text included in the prompt can travel, encrypted, to paired devices when you choose distributed inference.

Device hardware

MLXHub reads your device's RAM capacity to filter the model catalog and show RAM badges.

Settings

The app stores your preferences (port, Bonjour name, API key, system prompts, personality text) on-device only. These are never synced to iCloud or sent to any server.

Purchases

Payments are handled by Apple through the App Store. RevenueCat manages purchase history and entitlements under a random installation identifier, which the app also uses for analytics; this can associate purchases and diagnostics within an installation. The app does not give RevenueCat your name or email or use an advertising identifier. Apple Ads attribution collection is enabled in RevenueCat, so available attribution information can also be processed. Turning app analytics off does not stop the network operations needed to process or restore purchases. Apple's and RevenueCat's policies apply: apple.com/legal/privacy and revenuecat.com/privacy.

Native tools

When you ask for something in chat and approve it, MLXHub can call native tools that read — and for a few of them, write — data from system frameworks or from the app's own storage. These tools run on your device: what they read or write moves between the system framework, or the app's storage, and the local model. Two of them, described below, reach outside your device as a consequence of what you asked.

What each tool can do

Tools, including web search, grouped by the data they access:

  • Health: reads activity (steps, distances, energy), cardiovascular data (heart rate and its variability, blood pressure, oxygen saturation, respiratory rate, VO₂ max), sleep and wrist temperature during sleep, workouts, nutrition, body measurements, mindfulness sessions, state of mind and sexual activity; can log body measurements, nutrition, water, mindfulness sessions, state of mind (including the labels and associations you would record in the Health app) and sexual activity. Menstruation, pregnancy and other reproductive-health data are deliberately excluded.
  • Calendar: reads and can create events.
  • Reminders: reads and can create reminders.
  • Contacts: searches your contacts.
  • Location: reads your current location.
  • Motion: reads accelerometer, compass, and altitude data.
  • Music library: reads your library and controls playback. If you ask for a song you don't have, its title is looked up in the Apple Music catalog — see below.
  • Photos: searches your photo library and can attach photos to the chat.
  • Image reading: analyses an image you attached with Apple's on-device Vision framework — text recognition, barcodes, classification and the number of faces present. No face is identified and nothing is stored.
  • Memory: saves or forgets a fact about you — see “Assistant memory” below.
  • Conversation search: searches your own earlier conversations, stored in the app.
  • Skills: reads and writes files for the Skills you define, inside the app's own sandbox.
  • Text analysis, calculation, date & device info: work on what the model passes them or on basic device metadata — no user data involved.

You approve every call

Each tool call shows up as a card in the chat that you approve or reject individually — this is the default, most restrictive mode; you can pre-approve specific tools if you prefer. Tools that write data (Calendar, Reminders, Health, Memory) additionally require MLXHub Plus; that's a product decision, not a change to what data is read or sent. Every tool also needs the matching system permission (Health, Calendar, Contacts, Location, Motion, Music, Photos, Reminders) the first time it runs — without it, the tool simply fails.

Exception: web search

When you ask MLXHub to search the web, the text of your query leaves your device and goes to a third-party search engine — DuckDuckGo by default, or 360 Search if you choose it in Settings (needed where DuckDuckGo doesn't resolve). Nothing else about you or your device goes with it: no API key, no account, no persistent identifier, and no cookies kept between searches.

Exception: Apple Music catalog

When you ask to play a specific song that isn't in your library, MLXHub looks it up in the Apple Music catalog through Apple's MusicKit framework, which sends the title you asked for to Apple and checks whether you have an Apple Music subscription. Reading your library never does this. Apple's privacy policy applies: apple.com/legal/privacy

Assistant memory

MLXHub can remember facts about you — your name, what you're working on, how you like answers phrased — and carry them into later conversations by adding them to the model's instructions. There is a hard cap of 60 memories.

Who writes them

You do, in Settings → Memory, where you can add, edit, pin, delete one or delete all of them. The model can too, through the Memory tool, but only after you approve that specific call, and only with MLXHub Plus. The model cannot read the memory store separately: what it knows is already in its instructions.

Where they live

Memories are stored on your device, in the same database as your conversations. Like your conversations, they are not synced through iCloud, but they are part of your device backup (iCloud Backup or a local one) — they are not excluded from it. Each conversation also keeps a frozen copy of the memory block it started with, so deleting a memory in Settings does not rewrite older conversations that had already used it.

Where they never go

Memories are never included in analytics: the app only records that the memory tool was used, never what it saved or how many memories exist. They are never visible to clients of the LAN server, whose requests are built without them. The only path by which they leave your device is distributed inference, where they travel inside the prompt, encrypted, to the devices you paired.

Pseudonymous analytics & diagnostics

Analytics and diagnostics help understand app usage and find errors. They are enabled by default and can be turned off in Settings.

What is sent

The following usage and diagnostic events can be sent:

  • The app was opened — at most once per session.
  • Which features you use — chat, the LAN server, hosting a mesh, model slots, themes, or a tool that writes data — once per feature and session.
  • A model finished loading — whether it succeeded, roughly how long it took, a size range, and its kind (language, vision, embedding or reranking model).
  • A model finished downloading — whether it succeeded, and a size range.
  • An error or crash occurred — which part of the app it came from, and what kind of failure it was.
  • A generation was interrupted — and why (you cancelled it, memory ran out, or it failed).
  • The app freed memory after a critical memory warning from the system — with memory figures as ranges.
  • You called a native tool — which one (for example, reading Health data, saving a memory or searching your conversations), whether it succeeded, and roughly how long it took — never what you asked for, read, wrote or found.
  • A Skill was activated — where it came from, never its name or contents.
  • You rated an answer with a thumbs up or down — the rating and the model family, never the message.
  • You hit the MLXHub Plus paywall on a tool that writes data — which tool, nothing else.
  • Which screen you visited, onboarding progress and paywall outcomes — bounded categories, never conversation content.
  • An installation profile: preferred language, number of installed models, purchase-entitlement state, range of days since installation and whether distributed inference has been used.

App events include the app version, device class, chip family, RAM and GPU-memory ranges, and the configured device region. That region is a setting, not your physical location. Properties use predefined categories. Some measurements remain ranges; observed active durations of completed sessions can instead be sent as bounded totals rounded down to 1 second, with sample counts, to calculate observed totals and averages. Bounded event counts and indicators of exceeded limits or incomplete coverage can also be sent. Interrupted sessions do not invent missing time. Conversation content and arbitrary text are excluded.

Crash reports

If the app crashes, MLXHub captures a symbolized stack trace and the signal or runtime trap involved, using PLCrashReporter. The crash message itself — the text of a fatalError or an exception's reason — and the on-disk path of every loaded binary are stripped out before anything leaves your device; only the redacted trace is sent. This does not cover the system killing the app for using too much memory, which no crash reporter can see — that's counted separately as an interrupted generation, above.

What is never sent

Analytics never includes conversations, prompts, model answers, file paths or names, private/custom model repository names, tool inputs or results, images or audio, assistant memories, Skill names or rated-message text. It may include the name and publisher of a model only when they belong to the public catalogue shipped with the app; other repositories are grouped as “other”.

Persistent random identifier

Events are grouped under a random identifier generated on your device and retained across sessions, which allows usage and retention to be measured. It is not a name, email, Apple ID or advertising identifier. This is pseudonymous data, rather than a guarantee that events cannot be correlated. This identifier is also used for RevenueCat purchase entitlements and attribution, so analytics and purchases can be associated for the same installation; it is not an advertising identifier and is not used to track you across apps or websites.

Turning it off

Settings offers three levels, and the choice takes effect immediately:

  • Off — nothing is sent, and anything still queued on your device is deleted.
  • Anonymous — the default. The events above.
  • Extended — reserved for data that would need your explicit consent. Nothing currently uses it.

Choosing Off costs you nothing: no feature is disabled, degraded, or withheld.

Where it goes

This data is processed by PostHog, in their European data centre, for both analytics and crash/error reporting. PostHog's privacy policy applies: posthog.com/privacy

Model downloads, first use and performance

Downloads and loads can include model type, family, parameter range, public catalogue name and publisher. Failed or cancelled downloads include a bounded reason such as network, incompatibility, integrity or cancellation. The interval from an observed download to its first successful inference is sent as a range; local pending records expire after 90 days and are removed when the model is deleted or analytics is disabled. Performance events contain ranges for token counts, processing and generation timing or speed, and categorical context, cache, thermal and Neural Engine state. They distinguish enabling the Neural Engine from actually executing work on it; no prompt or generated text is collected.

Location inferred from the connection IP (GeoIP)

For new analytics events, PostHog can infer country, region, city and other approximate geographic attributes from the network IP address used to connect. The app does not request GPS or precise coordinates for this purpose. This processing is separate from the Location tool you explicitly approve, and from the configured device region. Older events are not backfilled with invented locations. Turning analytics off prevents these app analytics requests. The event also carries a bounded screen name; a compatibility alias for the same screen visit adds no conversation content.

When analytics is on, events can record onboarding starts, the assistant or developer-tools route, progress, completion and interruption, and public-catalogue model selection and download outcomes. Paywall events can include where it opened, the selected plan and a confirmed, cancelled, pending or failed purchase outcome. Server events can include startup outcomes, whether its website was enabled, page starts per server session and request counts by predefined endpoint family. Voice events can include activation, duration, completed-turn ranges and operational signals such as denied permissions, errors, retries, empty captures and early exits. These signals do not infer feelings; normal speech interruptions are not classified as frustration.

Website page-start counts use a random token that lasts only for that page and is checked locally; the token is not sent to analytics. These counts are page starts, not unique people or TCP connections. Analytics receives predefined endpoint families, not raw URLs, query strings, hostnames, API keys, browser fingerprints or LAN client IP addresses. Audio, transcripts, conversation text, prompts and generated content remain excluded. The connection-IP and approximate-location processing described below is separate from LAN client addresses.

App recommendations

If you have not made any purchase, Settings shows a small banner recommending other independent apps, through Kickstart Exchange. It disappears permanently as soon as you buy anything.

Kickstart Exchange matches apps, not people. It receives details about the app itself — bundle identifier, platform, app and build version, and App Store country — plus a receipt signed by Apple that is used only to verify the installation is genuine and is not stored. It does not receive any information about you, and no identifier of you or your device is involved. Kickstart Exchange's privacy notice applies: exchange.kickstart.tools/privacy

Privacy controls and limits

  • Analytics does not ask for your name or email. Its random identifier can connect events across sessions.
  • Does not use advertising identifiers (IDFA/IDFV) and does not track you across apps or websites.
  • System data is accessed only for actions you request and authorize, using the matching system permissions. This includes native tools and saving a generated image to Photos.
  • Does not send what a tool reads or writes, or what the assistant remembers, to any server of ours. The only ways anything you typed leaves your device are the four described above: a web search you request, an Apple Music catalog lookup you request, ModelScope downloads if you enable them, and distributed inference between your own paired devices.
  • The microphone is used only after system permission when you start voice chat. The camera is used when you choose to attach a photo.
  • Conversation content is never sent to our analytics providers. Explicitly requested web, music and paired-device actions have the network exceptions described above.
  • Local chat does not require an app account. An iCloud account is required only if you choose to publish or manage benchmark submissions.
  • Does not sell, rent, or share data with anyone beyond the providers named in this policy.

Third-party services

MLXHub makes external network requests to these services: Network requests reveal the connection IP address to the receiving provider, even when no name, email or app identifier is included.

  • Hugging Face (huggingface.co) — to download model files when you initiate a download: huggingface.co/privacy
  • Apple (App Store) — to process in-app purchase payments: apple.com/legal/privacy
  • Apple (Apple Music) — to look up a song you asked to play that isn't in your library: apple.com/legal/privacy
  • RevenueCat — to manage entitlements for in-app purchases: revenuecat.com/privacy
  • PostHog — pseudonymous analytics, crash reports and diagnostics processed in the European Union, unless disabled: posthog.com/privacy
  • Kickstart Exchange — to fetch app recommendations, shown only if you have made no purchase: exchange.kickstart.tools/privacy
  • DuckDuckGo (duckduckgo.com) — to run a web search you request; only the query text is sent: duckduckgo.com/privacy
  • 360 Search (so.com) — an alternate search engine you can select in Settings, used where DuckDuckGo doesn't resolve.
  • ModelScope (modelscope.cn) — an alternate model download source, active only if you turn it on in Settings.
  • Apple (CloudKit) — optional public benchmark publication linked to your iCloud account: apple.com/legal/privacy

Children

MLXHub does not knowingly seek personal information from children under 13. The analytics practices described in this policy apply.

Optional public benchmark ranking

Running a benchmark keeps its results on your device unless you explicitly publish them. Publication requires an iCloud account and uploads to Apple's public CloudKit database: device specifications, app and OS versions, benchmark models and settings, measured results and diagnostic evidence, and an optional country you declare or omit. The initial country comes from the device Region setting, not GPS. CloudKit links the record to its creator account, and the submission includes a random installation identifier to distinguish devices. This publication is linked to an account and is separate from pseudonymous analytics; turning analytics off does not delete a submission. No conversation, prompt, generated answer, audio or image is included. You can delete your submissions in the ranking's “My submissions” screen. Apple's privacy policy applies: apple.com/legal/privacy.

Voice and image generation

Voice chat uses the microphone after system permission and when you start it. Speech transcription and speech synthesis run on your device; language assets or neural-voice weights may need to be downloaded first. Transcripts become chat messages under the same local-storage and user-requested network rules as typed text. Image generation and its safety classification run locally after the necessary models are downloaded; saving an image to Photos is an action you choose and authorize. These features do not send audio, images, transcripts or generated content to analytics. Existing web-search, music-lookup and paired-device exceptions remain applicable when you request them.

Changes to this policy

This policy applies to the selected app version. It will be updated if that version's data practices change. The date above is the policy revision date.

Questions about this policy: open an issue at the project repository.