Privacy Policy

Last updated: November 16, 2026 · Effective date: October 30, 2026

In short: Pocket LLM Chat runs language models entirely on your device and has no accounts, no analytics, and no crash reporting. Your prompts, conversations, and model outputs are never transmitted to us.

The app makes network requests in exactly one situation: downloading model files. Those requests go to the model host you select in the app's HF Source menu — hf-mirror.com by default, or huggingface.co if you switch it. Whoever operates that host receives your IP address and standard request metadata, and is an independent party to us. Details in What Leaves Your Device and Model Hosts and Third-Party Recipients.

1. Overview

Pocket LLM Chat (shown as “Pocket LLM” on the Home Screen; formerly LocalLLMClient) is a local-first LLM chat app for iPhone. Model inference — the computation that turns your prompt into a response — runs on your device using Apple MLX.

Your privacy is our top priority. This policy explains what data the app handles, what leaves your device, and what we do with it.

2. Data We Collect

We do not collect any personal data. There is no server-side component to this app: we operate no accounts, no backend, and no user database, so there is nothing for us to hold.

  • All model inference runs entirely on your device.
  • No prompts, conversation history, images you attach, or model outputs leave your device.
  • No analytics, telemetry, or crash reporting is sent to any server.
  • No advertising or cross-app tracking.
  • No user accounts, authentication, or identity tracking.
  • We do not use third-party advertising or data-broker SDKs.

3. What Leaves Your Device

The app makes network requests for one purpose only: obtaining model weights from a model host. Once a model is downloaded, the app is fully functional offline and makes no further requests.

RequestWhenPurpose
GET /api/models/{repo}/revision/main Before a download List the model repository's files and sizes, so the app can show a download size and confirm the model exists.
HEAD /{repo}/resolve/main/{file} Before a download Read file size, ETag, and commit hash to skip files already present and resume partial downloads.
GET /{repo}/resolve/main/{file} During a download Download the model weight file itself.

Downloads may continue after you leave the app, using the operating system's background transfer facility. No other network activity occurs.

What the model host receives

Because the app must identify what to download, the request necessarily contains:

  • Your IP address, as visible to any web server you connect to.
  • Standard HTTP request metadata sent by the operating system's networking stack, such as the User-Agent string and requested file paths.
  • The repository and file names being requested — for example, LiquidAI/LFM3-1.7B-MLX-4bit. These are public model identifiers, chosen by us, not information about you.

What the model host never receives

  • Your prompts, conversations, or model outputs.
  • Images you attach to a message.
  • A Hugging Face account or access token — the app does not authenticate, and sends no credentials.
  • Any identifier we assign to you, because we do not create any.

Attached photos are read through the system photo picker, so the app never gets access to your photo library, and the image is processed on your device.

Offline use

After the models you want are downloaded, you can disable network access entirely — for example by turning off Wi-Fi and cellular data, or by putting the device in airplane mode. Chat, tool calling, and history all continue to work.

4. Model Hosts and Third-Party Recipients

This section is provided for transparency about regulatory obligations and data flows.

The app lets you choose which host serves model files, via the HF Source menu in the app's toolbar:

SelectionHost contactedWho operates it
Default hf-mirror.com A third-party mirror operator, not Hugging Face. It is a separate, independently operated service that proxies requests to model repositories.
Optional huggingface.co Hugging Face, the model hub operator.

Who acts as a data controller for downloads. The operator of the host you select receives your IP address and request metadata and acts as an independent data controller for that request. With the default setting, that party is the operator of hf-mirror.com, which is a third party and is not Hugging Face; if you switch the source to huggingface.co, it is Hugging Face. We do not control, own, or store any of this data, and we have no access to it. Those hosts process it under their own terms and privacy policies.

We do not send any personal data to a model host beyond what an ordinary file download requires. We do not use a model host as an analytics provider, and we have no visibility into what it logs.

5. Data Stored on Your Device

The app writes the following to your device, all inside its own sandboxed container. None of it is uploaded by the app.

WhatWhereNotes
Conversation history Application Support/LocalLLMClientExample/History/<model>-<id>.json and <model>-<id>.md One file per model, holding text messages only. The .md file is a human-readable copy of the same content.
Downloaded model weights Documents/.localllmclient/huggingface/models/{owner}/{repo}/ Public model files fetched from the selected host. The models offered here are about 1.0 GB (LFM3 1.7B) and 1.2 GB (Bonsai 2 2B). The app asks the system to exclude this directory from device backups.
Diagnostics log Documents/diagnostics.log A capped file (512 KB, then rotated) of timing and memory metrics, the active model identifier, and thermal state, used to diagnose app termination and streaming performance. It contains no prompt or response text and stays on your device unless you choose to share it. The app asks the system to exclude this file from device backups.
Exported conversations Location you choose when using Export… Written only at your request, as a Markdown file in the Files app or via a save panel.

Conversation history is retained until you delete it or uninstall the app.

Device backups are separate. This app does not transmit your data, but if you have iCloud or Finder/iTunes device backup enabled, your operating system may include app data in that device backup. That transfer is performed by Apple under your Apple account settings, is governed by Apple's privacy policy, and is controlled by your backup settings — not by this app. You can review and change it in your device's backup settings.

5.1 Shared Container With a Companion App

Pocket LLM Chat can work with a companion app from the same developer, installed separately on the same iPhone. The feature is optional and is used only when you trigger it. It relies on an App Group container (group.com.vega.personalassistant) that only these two apps can read.

  • Answering a one-off request. The companion app writes a request as JSON into that container and opens this app. The reply is generated here, on your device, written to the same container, and the companion app is reopened. The request and the reply are deleted as soon as they are read, and anything left behind is removed within 24 hours.
  • Sending a chat. Send chat to the companion app copies the current conversation into the container as a note; the companion app imports it and then deletes it.

Nothing in the shared container is uploaded, a one-off request is answered without touching your own conversation history, and if you never use either action no file is written there.

6. Deleting Your Data

You are in control, and deletion is immediate and local:

  • Clear Chat in the app's menu clears the current conversation from memory and deletes its stored history files for that model.
  • Removing a model's data deletes that model's downloaded weights and its conversation history.
  • Uninstalling the app removes its entire container, including history, model weights, and the diagnostics log.

Because we hold no data, there is nothing to request from us. Once data is deleted from your device, it cannot be recovered by us.

7. Third-Party Components

The app depends on the following open-source and platform components. All inference components run locally and none of them transmit your conversations:

  • MLX and mlx-swift — Apple's machine learning framework for Apple silicon.
  • See ⋯ → Licenses… in the app for the full list of bundled open-source dependencies and model license notices.

These are local software components; including a library does not send any data to its maintainers. No data is sent to them.

8. Your Rights

We do not collect data, so there is nothing for us to access, correct, export, or delete on our end. All of your data stays on your device, under your control, and can be deleted at any time as described above.

If you have a question about this policy or the app's data handling, contact us using the details in Contact.

9. Children's Privacy

Pocket LLM Chat is intended for a general audience and is not directed at children under 13. We do not knowingly collect personal information from children. Because the app collects no personal information from anyone, no special handling applies.

10. Security

Model inference and conversation data stay within the app's sandboxed container on your device. The app uses the operating system's TLS-protected networking for downloads and does not implement its own cryptography or store credentials. We recommend keeping your device protected with a passcode and current software.

11. Policy Updates

If we change this policy, we will update the "Last updated" date at the top of this page and, where appropriate, notify you in the app. Continuing to use the app after an update means the revised policy applies.

12. Contact

If you have questions or concerns about this privacy policy, please contact us:

Email: vegaodm@outlook.com