Guides · Updated
What leaves your PC? A privacy checklist for local AI
Which features in Ollama, LM Studio, Open WebUI and the voice tools send data to the internet, and the documented setting that turns each one off.
A model that runs on your computer keeps that chat on your machine. A few features in the apps these guides use are built to send something further: a cloud model, a web search, a realtime voice call, a version check, or a usage report. This page lists those features and the off switch their own docs name.
It covers Ollama, LM Studio, Open WebUI, and the voice tools from the speech guides: Whisper and the local programs built on it, Piper, Kokoro, Buzz, browser speech recognition, and Home Assistant’s Assist. The workflow for documents, the test with the network unplugged, and disk encryption are in Local AI for privacy, step by step.
Where a project’s docs do not name a switch, this page does not invent one. A model download is listed when the docs describe it. That request fetches a file onto your PC. It is separate from sending a prompt or a recording.
Ollama
Ollama’s FAQ says that a local run stays on your machine: they do not see your prompts or data when you run locally. Cloud models are the exception. In the app or the ollama command, a cloud model uses a :cloud tag, for example gemma4:cloud. The prompt for that model is processed on Ollama’s servers. Ollama’s cloud page says those prompts and responses are used to answer the request, are not used to train models, and, in their words, are not stored or logged. The same FAQ says the cloud service collects basic account info and limited usage metadata, and that this metadata does not include prompt or response content.12
Turn cloud models and Ollama web search off. The FAQ’s local-only mode disables both. Use either of these, then restart Ollama:
{
"disable_ollama_cloud": true
}
That file is ~/.ollama/server.json. On Windows that is %HOMEPATH%\.ollama\server.json, the same .ollama folder the Windows install guide uses for models.
Or set the environment variable, the same way the install guides set other Ollama variables:
OLLAMA_NO_CLOUD=1
After the restart, Ollama’s logs show Ollama cloud disabled: true. You lose cloud models and Ollama’s web search. That is the point of the switch.2
Ollama web search sends the query to Ollama. The web search API is POST https://ollama.com/api/web_search, and the body includes the query. It needs an Ollama API key. The cloud disable switch above turns this off with the cloud models.3
Calling https://ollama.com/api/chat with an API key also sends the prompt to Ollama, including when the model name in that request has no :cloud on it. The cloud page uses an OLLAMA_API_KEY for that. Leave that key unset.1
Leave the API on this PC. Ollama binds 127.0.0.1 port 11434 unless you change it. Set OLLAMA_HOST only if you mean to. 0.0.0.0 opens it to your network, and the API has no login of its own. To put it back, set OLLAMA_HOST=127.0.0.1:11434 and restart. The install guides show where that variable goes on Windows, Mac and Linux. Do not port-forward 11434.2
Phonon, Ollama’s speech model, is a local download. ollama pull phonon, then ollama run phonon, speaks on this PC through the local server. The speech docs do not describe a cloud voice. If the chat model in front of it is a :cloud tag, that prompt still goes to Ollama’s cloud before Phonon speaks the answer.4
LM Studio
LM Studio’s docs say it can run entirely offline once the model files are on the computer. Chatting with a downloaded model stays on the device, and a document you drop into the chat is processed locally.5
These actions make a network request, as that same page describes them:
- A search in the Discover tab, which contacts Hugging Face. The search text goes out. Skip Discover and sideload a model file you already have.
- Downloading a model, or loading stats and download options for the built-in catalog. Those requests go to Hugging Face.
- Checking for and downloading LM Runtimes, the engines such as llama.cpp.
- Opening the app on macOS or Windows. The built-in updater checks for a new version and can show an “update now or later” notice. The docs describe that check. They do not name a switch that turns the check off.
Keep the local server on this PC. The Developer tab can serve the API on localhost or on the local network. Serve on Local Network lets other devices on your network use that server. Leave it off. The docs’ own examples call a server on this PC at http://localhost:1234. lms server start --cors turns CORS on. The CLI docs say to leave CORS off unless you need it, because it widens who can call the server.6
MCP. LM Studio can call MCP servers during a chat. A server with a remote server_url receives the tool call. The docs’ example is https://huggingface.co/mcp. Two settings gate this, both in Server Settings: Allow per-request MCPs, and Allow calling servers from mcp.json. Leave both off when the chat should stay on the PC.7
The LM Studio pages cited here do not name an analytics checkbox. This page does not invent one.
Open WebUI
Open WebUI runs on your machine and talks to a local Ollama from there. The paths below are the ones in the current docs: Settings → Admin → …. On Docker, add -e flags to the command in the Open WebUI guide. For the pip install, set them in the environment before open-webui serve. Many of these are saved from the admin screen after the first run, so use that screen when you can see the switch.8
Web search
Web search sends the query to the provider you select. The docs’ provider list includes SearXNG, Brave, Google, Bing, DuckDuckGo, Tavily, Exa and others, including Ollama’s hosted web search. A provider you do not run yourself receives the query. A SearXNG address on your own machine receives it there, and that SearXNG then contacts the search engines it is configured to use.9
Turn it off. Settings → Admin → Web Search, and turn web search off. The environment variable is ENABLE_WEB_SEARCH=false. That is the default, so a fresh install is already off. There is a separate confirmation switch, ENABLE_WEB_SEARCH_CONFIRMATION, which asks before a query is sent. Confirmation is a prompt. The off switch is web search itself.
The default page reader stays in Open WebUI. A hosted loader such as Tavily or Firecrawl sends the page address to that service. Leave the loader at its default.
Cloud model connections
ENABLE_OPENAI_API defaults to on, which means the OpenAI-compatible connection is available. A chat that uses a model on that connection is sent to the saved base URL. The default base URL is https://api.openai.com/v1.
Turn it off. In Settings → Admin → Connections, remove the API keys and remote addresses, or set ENABLE_OPENAI_API=false. Leave the Ollama connection enabled. ENABLE_DIRECT_CONNECTIONS defaults to off. Leave it off: a direct connection lets the browser call that server itself.8
Call mode and speech engines
Admins choose the call mode in Settings → Admin → Audio → Voice calls.
Standard is the default. It uses the speech-to-text and text-to-speech engines. It stays on the PC when those engines are the local ones below and the selected chat model is a local model.
Realtime connects the call to a speech-to-speech model on an OpenAI-compatible Realtime API, or on Azure OpenAI. The default provider is OpenAI, and the default base URL is https://api.openai.com/v1. Your audio and the spoken replies pass through Open WebUI to that provider. At the start of the call the voice model also receives the visible text of the chat, up to the last 100 messages. Greetings and other small talk are answered by that voice model. Real questions are passed to the model selected in the chat, then the voice model speaks the answer.10
Turn Realtime off. Set Call mode to Standard, or set AUDIO_REALTIME_ENABLED=false. That is already the default. The docs also allow a Realtime base URL on a server you run yourself. For a setup that does not leave the PC, Standard plus the local engines is the documented path. Leave the Realtime base URL unset.
In Settings → Admin → Audio:
- Speech-to-text. An empty
AUDIO_STT_ENGINEis the default and runs local faster-whisper inside Open WebUI.openai,deepgram,azureandmistralsend the recording to those services. In your own Settings → Audio, set Speech-to-Text Engine to Default. Web API uses the browser’s speech recognition. On Chrome, that recognition sends the audio to a web service.1112 - Text-to-speech. An empty
AUDIO_TTS_ENGINEis the default. Playback then uses the browser’s Web Speech API, or the Browser Kokoro option in user settings.transformersruns a local model on the Open WebUI server.openai,openai-realtime,elevenlabs,azureandmistralsend the text to those services. Point an OpenAI-compatible voice URL at a Piper or Kokoro server on127.0.0.1when you want that local voice. - Whisper model updates.
WHISPER_MODEL_AUTO_UPDATEdefaults to false. Leave it false. The first use of local Whisper still downloads the model file into Open WebUI’s cache. Later transcription runs there.
Documents and images
Embeddings. The default engine is local SentenceTransformers (sentence-transformers/all-MiniLM-L6-v2), which downloads that model from Hugging Face once. ollama uses your local Ollama embedding model. openai and azure_openai send the chunks and the query to those services. Leave the engine empty, or set RAG_EMBEDDING_ENGINE=ollama.8
Content extraction. An empty CONTENT_EXTRACTION_ENGINE uses the built-in PyPDF loader inside Open WebUI. Mistral OCR, Azure Document Intelligence, and an external URL send the document to that service. Tika and Docling run wherever you point them. Use an address on this PC if you use those. The screen is Settings → Admin → Documents.13
Images. ENABLE_IMAGE_GENERATION defaults to false. If you turn it on, the default engine is openai, which sends the prompt to OpenAI’s images API. Set IMAGE_GENERATION_ENGINE to comfyui or automatic1111 and point it at a local install, or leave image generation off. The Open WebUI guide connects this to a local ComfyUI.
Version check, community sharing, telemetry
| Switch | Default | What the docs say it does | Set it to |
|---|---|---|---|
ENABLE_VERSION_UPDATE_CHECK | true on release builds | Checks whether a newer Open WebUI release exists | false |
OFFLINE_MODE | false | Forces the version check off, and stops automatic downloads of embedding, reranking and Whisper models | true only if those models are already downloaded |
ENABLE_COMMUNITY_SHARING | true | Shows Share buttons that upload the prompt, model, tool or chat you choose, and a stats sync that sends usage statistics to the Open WebUI community | false |
ENABLE_OTEL | false | Exports traces, metrics and logs to the OpenTelemetry address you set. The default address is http://localhost:4317 | false |
HF_HUB_OFFLINE | 0 | Stops Open WebUI connecting to Hugging Face for automatic model downloads | 1 after the files are already on disk |
HF_HUB_DISABLE_TELEMETRY | unset | Hugging Face’s own switch. 1 turns off telemetry from libraries that use the Hub | 1 |
OFFLINE_MODE does not turn off web search, and it does not turn off an OpenAI connection. Use the switches above for those. Community sharing does not upload every chat on its own. It uploads an item when someone uses Share, and the stats sync sends usage statistics. Turning ENABLE_COMMUNITY_SHARING off hides both.814
-e ENABLE_VERSION_UPDATE_CHECK=false
-e ENABLE_WEB_SEARCH=false
-e ENABLE_OPENAI_API=false
-e ENABLE_COMMUNITY_SHARING=false
-e ENABLE_OTEL=false
-e HF_HUB_DISABLE_TELEMETRY=1
OLLAMA_NO_CLOUD=1 belongs on the Ollama process, from the Ollama section above. HF_HUB_OFFLINE=1 is worth adding on Open WebUI after the Whisper and embedding files are downloaded. Set it earlier and the first document upload cannot fetch its embedding model.
Voice tools
Whisper, whisper.cpp and faster-whisper
| Tool | What the docs say touches the network | How to keep the recording local |
|---|---|---|
| Whisper (the official Python program) | The README shows a local command, whisper audio.flac --model turbo. It does not document a cloud mode or a telemetry switch. | Run that local command. The recording is an argument to the program on this PC.15 |
| whisper.cpp | You download a ggml model with download-ggml-model.sh. The project describes on-device transcription. Its server example uses whisper-server --host 0.0.0.0, which listens on every interface. | Run the local whisper-cli. If you start the server, use --host 127.0.0.1 and do not publish the port.16 |
| faster-whisper | Loading a size such as large-v3 downloads that model from the Hugging Face Hub. | Transcription then runs locally. Set HF_HUB_DISABLE_TELEMETRY=1. Set HF_HUB_OFFLINE=1 only after the model file is cached.1714 |
Open WebUI’s local Whisper engine is faster-whisper, so the same two Hugging Face variables belong on that process.
Piper and Kokoro
Piper’s README describes a local text-to-speech engine, installed with pip install piper-tts. It also documents a web server. If you run that server for Open WebUI, bind it to 127.0.0.1.18
Kokoro’s README is a local inference library. The weights are the Hugging Face model hexgrad/Kokoro-82M, so the download contacts Hugging Face. The synthesis in the README runs in the Python process on this computer. The same Hub variables apply when the download goes through the Hugging Face libraries.19
Buzz
Buzz’s README is a desktop app you install and run on the computer. It does not name a telemetry switch or a cloud transcription setting. Use the local install. This page does not add a checkbox the README does not list.20
Subtitle Edit is in the speech guide as a local way to build subtitles with Whisper. Nothing cited here names a telemetry switch for it. Keep its transcription on the local Whisper engine.
Browser speech recognition is the Open WebUI Web API option above. Chrome sends that audio to a recognition service.12
Home Assistant
Home Assistant’s voice page says Assist can run on your own hardware, or through Home Assistant Cloud. Choose I plan to process my voice locally. A Cloud voice setup sends the audio to Home Assistant Cloud. Point the conversation agent at the Ollama integration on this PC, with Ollama still on localhost and cloud features disabled.21
Home Assistant analytics are separate from the voice pipeline. Nothing is sent unless you opt in. The sharing switches are at Settings → System → Analytics. Turn the options off there. When they are on, the report goes to analytics.home-assistant.io.22
Quick checklist
- Ollama:
OLLAMA_NO_CLOUD=1or"disable_ollama_cloud": true, then restart. NoOLLAMA_API_KEY.OLLAMA_HOSTleft on127.0.0.1:11434. - LM Studio: sideload models when you want to skip Discover. Serve on Local Network off. CORS off. Both MCP allow-switches off.
- Open WebUI: web search off, OpenAI connection off, call mode Standard, speech-to-text Default (local Whisper), text-to-speech on Browser Kokoro,
transformers, or a Piper or Kokoro server on localhost. - Open WebUI documents: embedding engine empty or
ollama. Content extraction left empty. Image generation off, or pointed at a local ComfyUI. - Open WebUI environment:
ENABLE_VERSION_UPDATE_CHECK=false,ENABLE_COMMUNITY_SHARING=false,ENABLE_OTEL=false,ENABLE_WEB_SEARCH=false,ENABLE_OPENAI_API=false,HF_HUB_DISABLE_TELEMETRY=1. - Voice: local Whisper, Piper or Kokoro. whisper.cpp’s server on
127.0.0.1. Home Assistant on the local voice path, analytics off.
Then prove the chat still answers with the network unplugged. That test is step 7 of the privacy walkthrough.
Sources
- Ollama, Cloud. In the app or CLI a cloud model is a
:cloudtag, for examplegemma4:cloud. Prompts and responses for cloud models are processed to answer the request. Ollama says they are not used to train models. Requests tohttps://ollama.com/api/chatuse anOLLAMA_API_KEYand send the prompt to Ollama. Ollama cloud. - Ollama FAQ. Local runs: "We don't see your prompts or data when you run locally." Cloud models: prompts and responses are processed to provide the service, and Ollama says they are not stored, logged, or used for training. The service collects basic account info and limited usage metadata that does not include prompt or response content. Disable cloud features with
disable_ollama_cloudset totruein~/.ollama/server.json, orOLLAMA_NO_CLOUD=1. Restart. The log line isOllama cloud disabled: true. That mode turns off cloud models and web search. The server binds127.0.0.1port11434unlessOLLAMA_HOSTchanges it. Ollama FAQ: disable cloud features. - Ollama, Web search.
POST https://ollama.com/api/web_searchwith aquerystring. Requires an API key. Ollama web search. - Ollama, Speech.
ollama pull phonondownloads the speech model.ollama run phononspeaks through the local server. The speech API examples usehttp://localhost:11434. Ollama speech. - LM Studio, Offline Operation, on the app docs. Once a model is downloaded, chat stays on the device, and documents dropped into chat are processed locally. Discover search, catalog stats, model downloads, runtime downloads, and the macOS and Windows update check on launch are the operations that use the network. The update check has no off switch on that page. Sideloading is the documented way to use a model file you already have. LM Studio app docs. Sideload a model.
- LM Studio, local server. The Developer tab serves on localhost or on the network. Serve on Local Network lets other devices on the network use the server.
lms server startstarts it.--portchanges the port. The quickstart uses port1234on localhost.lms server start --corsenables CORS, which the CLI reference says to leave off unless you need it. LM Studio local server. - LM Studio, MCP via API. A remote
server_urlreceives tool calls. The example URL ishttps://huggingface.co/mcp. Per-request servers need Allow per-request MCPs in Server Settings. Servers inmcp.jsonneed Allow calling servers from mcp.json. LM Studio MCP. - Open WebUI, environment configuration.
ENABLE_WEB_SEARCHdefaults to false.ENABLE_OPENAI_APIdefaults to true and the default base URL ishttps://api.openai.com/v1.ENABLE_DIRECT_CONNECTIONSdefaults to false.AUDIO_STT_ENGINEempty runs local faster-whisper.AUDIO_TTS_ENGINEempty uses the browser, including Browser Kokoro.AUDIO_REALTIME_ENABLEDdefaults to false (Standard).WHISPER_MODEL_AUTO_UPDATEdefaults to false.RAG_EMBEDDING_ENGINEempty uses local SentenceTransformers.CONTENT_EXTRACTION_ENGINEempty uses the built-in loader.ENABLE_IMAGE_GENERATIONdefaults to false, and the image engine defaults toopenaiif you enable it.ENABLE_VERSION_UPDATE_CHECKdefaults to true on release builds.OFFLINE_MODEforces that check off and stops automatic Hugging Face model downloads. It does not disable web search or external LLM APIs.ENABLE_COMMUNITY_SHARINGdefaults to true and covers Share actions plus usage-stats sync.ENABLE_OTELdefaults to false. The default OTLP address ishttp://localhost:4317.HF_HUB_OFFLINE=1stops Hugging Face downloads. Open WebUI environment configuration. - Open WebUI, Web search. The query goes to the search engine you configure under Settings → Admin → Web Search.
ENABLE_WEB_SEARCH_CONFIRMATIONasks before a query is sent. The default loader reads pages in Open WebUI. Hosted engines are alternatives. Open WebUI web search. - Open WebUI, Voice mode. Settings → Admin → Audio → Voice calls. Standard uses the STT and TTS engines. Realtime connects to an OpenAI-compatible Realtime API or Azure OpenAI. Audio and spoken replies pass through Open WebUI to that provider. The voice model receives the visible chat text. The default Realtime base URL is
https://api.openai.com/v1. A compatible server you run yourself is also accepted. Open WebUI voice mode. - Open WebUI, Configure speech-to-text. Local Whisper is the default. OpenAI, Deepgram, Azure, Mistral and a self-hosted OpenAI-compatible server are the other engines. Admin path: Settings → Admin → Audio. In user settings, Speech-to-Text Engine is Default or Web API. The Web API uses the browser's speech recognition. Open WebUI speech-to-text.
- MDN,
SpeechRecognition. On some browsers, including Chrome, recognition uses a server-side engine, so the audio is sent to a web service and does not work offline. MDN SpeechRecognition. - Open WebUI, document extraction and the environment reference. The default PDF loader is the built-in PyPDF loader, used when
CONTENT_EXTRACTION_ENGINEis empty. Mistral OCR, Azure Document Intelligence, and theexternalengine send the file to the service you configure. Admin path: Settings → Admin → Documents. Open WebUI document extraction. - Hugging Face Hub, environment variables.
HF_HUB_DISABLE_TELEMETRY=1disables telemetry from Hub libraries.HF_HUB_OFFLINE=1stops HTTP calls to the Hub, including the check for a newer cached file. Hugging Face Hub environment variables. - OpenAI Whisper README. Transcription is a local command:
whisper audio.flac --model turbo. The README does not document a telemetry switch or a cloud transcription mode. Whisper on GitHub. - whisper.cpp README. Models are downloaded with
download-ggml-model.sh. The project describes offline on-device transcription. The server example iswhisper-server --host 0.0.0.0. whisper.cpp. - faster-whisper README.
WhisperModel("large-v3")downloads the model from the Hugging Face Hub. Open WebUI's environment docs say its local Whisper backend is faster-whisper. faster-whisper. - Piper README. "A fast and local neural text-to-speech engine." Install with
pip install piper-tts. The project also documents a web server. Piper. - Kokoro README. A local inference library for the Hugging Face weights
hexgrad/Kokoro-82M. Kokoro. Kokoro-82M on Hugging Face. - Buzz README. Install and run the desktop app locally. The README does not name a telemetry switch or a cloud transcription setting. Buzz.
- Home Assistant, Assist. Assist can run on your own hardware, or through Home Assistant Cloud. The page links a local setup ("I plan to process my voice locally") and a Cloud setup. The Ollama conversation agent uses the Ollama server you configure. Home Assistant voice. Ollama integration.
- Home Assistant, analytics. Sharing is opt-in. Nothing is sent unless you turn an option on. Change or turn the options off at Settings → System → Analytics. Reports go to
analytics.home-assistant.io. Home Assistant analytics.
Get the Local AI weekly
New models worth running, what fits which GPU, and setup fixes. One short email a week, free. Unsubscribe any time.
Local-LLM App Starter
A tested starter for building your own app on Ollama: chat, RAG and a GPU-fit checker. Skip the setup hours.
kits.lilaforge.com → VibeEngineBest local LLM tools, ranked
Ollama, LM Studio, Open WebUI and llama.cpp scored from public evidence, with top picks.
vibeengine.lilaforge.com →