Apna is Mira — an AI agent that picks up your calls, replies to texts, and remembers the things you’d rather not. She runs on your phone, never sends your conversations to a server, and forgets things when you ask her to.
100% on-device · GDPR + EU AI Act compliant · Open source
CALLKIT · ON-DEVICE
Built withGemma 4Apple Foundation ModelsWhisperKitLiveKitCloudflare WorkersLadybugDB
01 · Calls
Don’t miss another call. Mira picks up.
When you can’t get to your phone, Mira picks up and takes a message. She transcribes the conversation in real time, summarizes it, and texts you the highlights. Everything runs on your phone — Apple’s servers never hear a word.
Real-time transcript
WhisperKit on iOS, Moonshine on Android.
In-app summary
Sent the second the call ends.
Optional callback
Mira can ring them back in your voice.
Zero data sent to a server
Speech recognition is 100% on-device.
02 · Model
A 4-billion-parameter model. Running in your pocket.
Apna uses Gemma 4 for general reasoning and Apple’s Foundation Models on iOS 26+ for tighter integration. The model lives entirely on your phone. Inference takes under 100 ms with no network round-trip.
4B
Parameters
0ms
Cloud latency
0 KB
Sent per query
100%
Offline-capable
mira · on-device shell
→ mira.ask("what did Sarah say?")
✓ matched 1 episode (2026-05-25, 14:32)→"Sarah from the bakery called. Your order is ready."_▌
03 · Voice
Mira can speak in your voice. Only if you want.
Train Personal Voice in five minutes, on your phone, overnight. Mira can then answer calls and reply with your voice. Voice cloning is opt-in, never shared, never used for anyone but you.
iOS 17+ · Apple Personal Voice entitlement AudioSeal frame-level watermark on all synthesized output
04 · Memory
Mira remembers everyone you talk to.
Every call, message, and event becomes a node in a graph that lives on your phone. Search it semantically, view it visually in Memory Inspector, export it as JSON or Markdown — and delete any node, any time.
When you message another Apna user, our server can’t read it.
Apna’s AgentLink protocol uses XChaCha20-Poly1305 with Ed25519 device keys. Your message is encrypted on your phone, routed as ciphertext through Cloudflare Workers, and decrypted only on the recipient’s phone.
iPhone ASENDERENCRYPTciphertext128B blobCloudflare Workerapna.workers.devSEES ONLY BYTESciphertextDECRYPTiPhone BRECIPIENT
Download Apna for iOS or Android. Onboarding takes 60 seconds.
2
Step 2
(Optional) Train
Record five voice prompts to teach Mira your voice. Trains overnight, on-device.
3
Step 3
Mira answers
She picks up incoming calls when you can’t. You stay in control.
Why Apna
Cloud AI assistants vs Apna
Cloud AI (ChatGPT/Gemini)
Apna
Where your conversations go
Server in another country
Your phone. That’s it.
Works without internet
No
Yes, after one-time model download
Answers your phone for you
No
Yes, on-device
Personal voice clone
Cloud-trained, shared infra
On-device, opt-in, yours alone
06 · Open
Built in the open.
Apna is a real-codebase product, not a wrapper. The agent loop, the protocol, the on-device model integration, the messaging server — all source-available. Read the architecture, propose changes, fork the parts you want.
No. Voice frames never leave your device. The on-device LLM (Gemma 4) handles transcription, reasoning, and replies locally. The only thing our server sees is opaque ciphertext for messages to other Apna users.
Free tier covers everything that runs on your phone. Pro adds outbound PSTN minutes (so Mira can call people back) and premium voice features. Pricing announced at launch.
Yes, after a one-time model download (~74 MB for Whisper, 2.1 GB for TTS). After that, the entire flow — call answering, transcription, replies — works on airplane mode.
Yes, with opt-in Apple Personal Voice (iOS 17+). Training takes about five minutes of recorded phrases and runs overnight on your device. Your voice model never leaves your phone.
The codebase is source-available today and will be public-license under MIT shortly. Build on it, fork it, replace any component.
Apna is built privacy-by-default. We comply with GDPR Articles 15-22 (access, erasure, portability) and EU AI Act Article 50 (transparency for AI-generated audio — every synthesized voice frame carries a watermark). Full details on the EU AI Act disclosure page.
Through Telnyx PSTN with your registered number as the caller ID. Outbound minutes are part of the Pro tier.
iOS 17 or later, and Android 15 (API 35) or later.