Your messages, documents, and personal files are processed by a language model running on your phone—but how does the app actually protect your data? This is a technical breakdown of MyBenAI's on-device architecture: the runtime, storage layer, optional network calls, and the network log that lets you verify what leaves your device.
Why Architecture Matters for Privacy
Privacy isn't just a feature you enable in settings. It's baked into the system at the architectural level. A cloud AI app can claim "end-to-end encryption" while still sending metadata, analytics, or model inputs to a remote server. A truly private app eliminates the possibility of that leakage by processing everything locally.
MyBenAI's architecture is built around a single principle: your message content is processed only on your device. No cloud LLM, no remote inference, no server-side database holds your chat history. Understanding how this works requires looking at three layers: the runtime environment, the storage layer, and the optional network boundary.
The Runtime: llama.rn and On-Device Inference
The heart of MyBenAI is llama.rn, a binding of llama.cpp for React Native. This is the inference engine that runs your language model locally. When you hit send on a message, the text stays in process memory. It's tokenized, passed to the LLM, and decoded into a response—all on the same device, using the app's CPU/GPU/NPU.
This is different from cloud AI, which sends your text over HTTP to a remote server. With llama.rn, there's no network step for chat inference. The architecture makes it physically impossible for your message content to reach a server, because the inference pipeline doesn't have a network call in it.
The same principle applies to embeddings and vision. When you ask MyBenAI to search your documents or analyze a screenshot, embeddings are generated by an embedding model (E5-small, Nomic Embed, or EmbeddingGemma) running locally via llama.rn. Vision processing uses SmolVLM or Qwen3-VL, also running locally. No vision API calls, no cloud image processing—the vision model and your image never leave your device.
Storage: SQLite, MMKV, and the Local Database
Data at rest lives in two places on your device:
- Chat and documents: Stored in SQLite via op-sqlite, the same database engine used by countless Android and iOS apps. Chat history is in plain SQLite—not separately encrypted, but protected by your device's full-disk encryption while the device is locked. Vector embeddings (for RAG) are stored in SQLite with the sqlite-vec extension. Full-text search indexes use FTS5. All of this is app-private storage, not readable by other apps.
- Secrets and settings: API keys (for optional web search), connector tokens, and encrypted preferences are stored in MMKV, a fast key-value store with encryption built in. Even if someone extracted your app's database files, they'd find encrypted tokens, not plain text.
The critical point: your chat database is not separately encrypted with a cipher like SQLCipher. It relies on the operating system's full-disk encryption. This is a trade-off—OS-level encryption is sufficient for phones that are typically locked and in your pocket, and it's faster than per-app encryption. But it means the database is decrypted when your device is unlocked. If you're concerned about forensic access after device theft, this is worth understanding.
Model Downloads and Hugging Face Integration
Language models are large—0.6B to 8B parameters. They're not bundled in the app (that would be 10+ GB). Instead, MyBenAI downloads models from Hugging Face at runtime, the first time you launch the app or when you change models.
This download is user-consented—you explicitly choose which model to download, and the app shows you the file size before downloading. The model file is downloaded directly from Hugging Face to your device storage, then loaded into memory for inference. The download itself is a legitimate network call, and it's one of the few network activities that MyBenAI performs for core functionality.
Once downloaded, the model file lives in your app's document directory. You can manually delete it if you want to free space, or delete it at app uninstall. MyBenAI never auto-updates models; you control which version runs.
Optional Network Calls: Web Search, Connectors, and Analytics
MyBenAI is designed to work offline in airplane mode. Your chat, documents, and calculations all work without internet. But the app has optional network calls you should understand:
- Web search: If you enable web search, the app queries an external search engine (Serper API if you provide a key, or DuckDuckGo as a free fallback). Your search query is sent to that service. The results are cached locally for 30 days to reduce repeated requests. Your chat prompts are not sent to the search service—only your search queries. This is user-controlled: you can disable web search entirely.
- Connectors: You can add remote MCP servers or REST API connectors (e.g., Home Assistant, OpenWeatherMap). When you use a connector, the app sends requests to that server. These connectors are optional and require explicit setup. MyBenAI doesn't force any connectors on you.
- Anonymous usage analytics: MyBenAI sends anonymous telemetry ON BY DEFAULT. This includes: your device RAM tier, which model you selected, token counts per chat, which personas you use, whether you've enabled RAG or long-term memory. It does NOT include message content, chat history, or document text. The data is sent to mybenai.robomiri.com. You can disable analytics in settings.
- Error reporting: Crashes and errors are sent to Sentry for diagnostics. This includes stack traces and device info, not your messages or documents. You can opt out.
The key insight: MyBenAI separates message content (which stays local) from operational telemetry (which is anonymous). This is different from "zero telemetry" claims, but it's honest: we track that you're using the app and which features you use, to understand product health. We don't sell or share that data.
The Network Log: Verify What Leaves Your Device
Trust, but verify. MyBenAI ships a built-in network log that shows you every HTTP request the app makes. You can view it in the admin console under Settings. The log lists the URL, headers, method, and response status for every request—model downloads, search API calls, connector requests, and analytics pings.
This is rare in the mobile AI app world. Most cloud-based AI apps don't give you visibility into what they transmit. MyBenAI's network log is an explicit commitment to transparency: if you're skeptical that your messages are really staying local, you can audit it yourself. Open the log, start a chat, and confirm that chat inference produces zero outbound requests.
Device-Tier Model Selection and Performance
MyBenAI automatically profiles your device's available RAM and selects the largest model that fits. This is baked into the architecture because model selection is critical for privacy: larger models (4B–8B) are more capable and leak less information through "hallucinations" or vague answers. A device with 12 GB RAM gets Qwen3.5 4B or Gemma 3 4B. A device with 4 GB RAM gets Qwen3 0.6B, which is smaller but still useful for local inference.
The selection happens silently at install time. You don't need to manually pick a model unless you want to override the default. This keeps the privacy boundary simple: one model, running locally, no ambiguity about where your data is processed.
Offline-First Design
The entire architecture is offline-first. Chat, search, documents, memory, tools—all of these work in airplane mode. Network access is opt-in: you enable web search if you want it, you add connectors if you need them. The default state is fully functional offline.
This is an architectural choice with a trade-off: without web search, MyBenAI's knowledge is limited to the model's training data, which has a cutoff date. A small model trained on 2024 data won't know about events in 2026. Cloud AI has live knowledge. But the privacy guarantee is absolute—if you never enable web search, nothing you say leaves your device.
One-Tap Erase and Data Deletion
If you decide to delete all your data, MyBenAI has a one-tap erase in Settings. This deletes your entire chat history, documents, RAG collection, long-term memory, downloaded models, and cached web search results. It doesn't uninstall the app, but it clears everything the app has learned about you.
Additionally, uninstalling MyBenAI removes all traces of its data from your device. iOS and Android enforce app-private storage isolation, so no MyBenAI data persists after uninstall.
No Account, No Login, No Cloud Sync
MyBenAI is a one-time $2 purchase with no account creation, no login, and no cloud sync. This simplicity is intentional: if there's no account server, there's no database of your identity linked to your chats. You're not a user ID in a system; you're an anonymous person running an app on your phone.
The trade-off: you can't sync your chats to another device. If you switch phones, your chat history doesn't transfer. You can export your data as JSON for backup, and restore it manually, but there's no seamless cloud sync. This is a deliberate privacy choice—the convenience of syncing isn't worth the infrastructure and the trust model it requires.
What This Architecture Enables
This on-device architecture enables a specific kind of AI assistant: one that's private by default, offline-capable, owned entirely by you, and auditable (via the network log). It's not the fastest AI (your phone is slower than a data center), and it's not the most capable (a 2B model can't reason like a 70B model). But it's genuinely yours, and nothing in its design creates a back door for data leakage.
To understand the broader landscape, explore what on-device AI is and how it compares to cloud AI. For a head-to-head privacy comparison, read cloud AI vs on-device AI. And to see exactly what network requests happen (or don't), check the network log guide. Ready to use an AI assistant that's truly under your control? Get MyBenAI for $2 and start using private AI today.