Privacy and where your data goes

Vanta stores app data on your device. Where a request is processed depends on the model, speech engine and tools you choose. A local chat model does not make an external search, embedding service or MCP tool local.

Model requests

Mode	Processing location	What is sent
Downloaded GGUF or MLX	Your iPhone or iPad	Chat inference stays on the device. Model discovery and downloads contact the model host.
Apple Intelligence	Apple’s on-device system runtime	Chat input is passed to the system model on your device.
Self-hosted server	The computer or server at your endpoint	The assembled request goes directly to that endpoint.
Cloud API	The provider at your endpoint	The assembled request goes directly to the provider.

A remote request can include conversation history, the system prompt, relevant memories, document passages, compatible images and tool results. Background work such as titles, summaries or memory extraction can also use a configured model endpoint. There is no Vanta AI server relaying those model requests.

Your chosen server or provider controls its own logging, retention and processing. Review its privacy policy. An HTTP connection on a LAN is not encrypted by HTTPS; use a trusted network, VPN or secured endpoint as appropriate.

Storage and keys

Conversations, presets, memories and document data are stored locally. Vanta’s app database does not enable CloudKit sync. API keys and MCP authentication secrets use the device Keychain; provider keys are kept out of the conversation database.

Files you export or share leave the app through the destination you choose. Device backups and any separately configured services are subject to their own settings and policies.

Files, PDFs and images

PDF text extraction and retrieval preparation happen in the app. If you ask a remote model about a PDF, the document text or retrieved passages needed for the request are sent to that endpoint. If you configure a remote embedding provider, text used to build embeddings is also sent to that provider.

Images are included only for a supported image workflow. GGUF and Apple Intelligence do not receive image input in Vanta. Selected MLX models process images locally; a remote vision model receives the image as part of the request.

Memory and embeddings

Memories are stored locally, but generating a memory or using it in a remote conversation may send the relevant text to your configured model. Automatic memory extraction and retrieval settings are separate controls.

Embedding settings are separate from chat settings. Choosing a local chat model does not change a configured remote embedding endpoint. Without an embedding model, memory retrieval can use keywords.

Speech

WhisperKit processes voice input on-device after its model is downloaded. Vanta’s Apple Speech path requires on-device recognition. Optional cloud transcription sends recorded audio to the configured transcription provider.

Search and MCP tools

Built-in web search sends queries to Tavily when you configure and use it. Research can make further online requests. MCP tool arguments are sent to the connected server, which may contact additional services or perform actions. Tool results may then be included in a model request.

Use servers you trust and review which tools are enabled. The MCP guide explains execution and access.

Collection and website analytics

Vanta’s privacy manifest declares no tracking, no tracking domains and no collected data types. That declaration does not remove the data flows to services you connect yourself.

This website may use Umami to measure page visits and App Store link clicks. Website analytics does not receive your Vanta conversations or API keys. Its hosting provider may handle standard request logs. Opening the App Store takes you to Apple’s service.

Contact

For privacy questions, email ezel964@icloud.com. Include only the information needed to answer your question; never send API keys or authentication tokens.