Your personal AI.
On your phone.
Not in the cloud.
Apna is an agent that lives on your device — answering your calls, taking messages, managing your time. The model runs locally. Messages between agents are end-to-end encrypted. Your memory never leaves the phone.
An agent that's actually yours.
Four engineering decisions add up to a product with a different shape than the assistants you're used to.
- On-device LLM
The model is in your pocket.
Gemma 4 runs locally. On iOS, Foundation Models is the fallback. Your conversations never leave the phone. No prompts shipped to a server. No telemetry on what you ask.
- Voice calls — answered
Mira picks up when you can't.
Incoming PSTN calls are answered by an AI agent that knows your context. Real-time ASR (WhisperKit, Moonshine) and TTS (TTSKit, Piper-VITS). Every synthesized voice is watermarked per EU AI Act Article 50.
- End-to-end encrypted messaging
Servers see ciphertext. Nothing else.
A2A protocol on top of Cloudflare Workers. Your agent talks to other agents through encrypted envelopes. The relay can't read them. We can't read them. We'd have to redesign the system to try.
- Memory you own
A local graph database, exportable as a ZIP.
LadybugDB on-device graph stores your memory, contacts, and conversation history. Optional Personal Voice clone. Full export in standard formats — no walled garden, no lock-in.
The compliance work is upstream of the product.
Apna is on-device-first because that's the architecture, not a marketing choice. Everything that follows — GDPR posture, EU AI Act disclosures, no tracking — falls out of that.
- GDPR (EU)
- EU AI Act, Art. 50
- AudioSeal watermarked TTS
- On-device by default
- No tracking, no analytics
Apna answers your phone.
Join the TestFlight to try Mira on iOS. Android beta opens alongside.