Your AI. Your phone.
Nobody else's business.
Dhruva is named for ध्रुव — the pole star, the one fixed point in a turning sky that never needs a network or an audience to hold steady. The app is that idea in your pocket: every model runs on your phone, every answer is generated offline, and the only network call it ever makes is the one you start yourself.
- Zero telemetry
- Apache-2.0
- Runs offline
- Any GGUF
Why Dhruva
Made for the moments cloud AI can't reach.
Every other AI app stops working the second the signal drops or the free tier runs out. Dhruva doesn't — because the model is already on your phone.
Offline, anywhere
On a flight, a trek, or abroad with no data — your assistant still answers. Turn on airplane mode and it doesn't even notice.
Private by physics
Ask about a photo, a contract, a medical report — the file never leaves your device. Not encrypted-in-transit; simply never sent.
No account, no bill
No login, no subscription, no per-message cost. Download a model once and it's yours to run forever, as much as you like.
See it, say it, offline
Ask about a screenshot or a photo, or talk hands-free while your hands are busy — vision and voice both run on-device, nothing uploaded. (Document chat is on the roadmap.)
What we're building
One app. Every model. Nothing leaves your device.
Dhruva is in active, open development. Here's the product taking shape, screen by screen — each preview below is drawn straight from the design system, not a marketing render.
A chat that never asks for a signal
Talk to any model you've downloaded, fully offline. Streaming responses with a live tokens-per-second readout, code blocks, and collapsible reasoning so you can watch the model think.
- Live tok/s ticker and markdown + code rendering
- Collapsible <think> reasoning, on by request not by default
- Regenerate, edit, export — history stays on device
final stack = <int>[]; stack.add(1); // push stack.removeLast(); // pop
Any GGUF from Hugging Face, no walled garden
Browse and pull quantized models straight from the Hub. Before you spend the bandwidth, Dhruva gives an honest device verdict based on the file size against your phone's RAM.
- Comfortable / Possible / Not recommended — no guessing
- Resumable, integrity-checked downloads
- Sideload your own fine-tunes too
Personas with their own mind and memory
Give a model a name, a voice, and a standing system prompt. Build a Socratic tutor, a dungeon master that keeps the campaign log, or a focused assistant — each with its own memory.
- Custom system prompts per character
- Separate memory, separate conversations
- For roleplay or narrow, reliable helpers
Point it at anything and ask
Show Dhruva a photo, a screenshot, a boarding pass. A vision model reads it and answers your questions about what's in the frame — image understanding that never leaves the phone.
- On-device image understanding, no upload
- Screenshots, documents, photos
- The mmproj loads alongside the base model
Make images without making a request
Quantized Stable Diffusion, generating on your phone's own compute. Type a prompt, watch it denoise step by step, and keep every pixel local — not one request leaves the device.
- Quantized diffusion, fully offline
- Live step-by-step progress
- Your prompts never touch a server
A real conversation, not a walkie-talkie
Speech in, speech out, with genuine turn-taking — the model listens, knows when you've finished, and can be interrupted mid-answer. Speech recognition and synthesis both run on-device.
- VAD, turn-taking, and interruption
- STT and TTS on-device
- No push-to-talk bolt-on
Drop in a PDF, ask it anything
Dhruva indexes your document with on-device embeddings, retrieves the relevant passages, and answers with a citation back to the page — retrieval-augmented, with nothing uploaded.
- On-device embeddings and retrieval
- Answers cite the page they came from
- Your documents stay yours
Small local tools, reached for mid-chat
Alarms, notes, quick calculations — the model can call small on-device tools right inside a conversation. The intent is handled on your phone; no assistant server is ever in the loop.
- Set alarms, save notes, run calculations
- Tool calls render inline and honestly
- Local intents, no cloud round-trip
Compare models. Tune them. Stay current.
Run one prompt across two models side by side, tune the sampling live, and read an opt-in digest of what's new in on-device AI — fetched only when you ask, never in the background.
- A / B compare with live sampling controls
- Opt-in AI news, fetched on request
Privacy, by construction
Not a policy.
An architecture.
Most "private" AI apps mean "we promise not to look." Dhruva means the data physically never leaves. There is no server in the loop to make a promise about.
Zero telemetry
No analytics SDKs, no crash reporters phoning home, no feature flags that call out. Nothing about how you use Dhruva leaves your device.
Downloads-only network
The single network call Dhruva ever makes is the one you ask for: fetching a model from Hugging Face. That's the entire network surface.
Open source, Apache-2.0
Every line is public. Don't take our word for the privacy claims — read the code, or point a network monitor at the app and watch.
The pole star doesn't rush to prove it's there.
Neither does Dhruva. Follow along, read the code, and help shape an AI that answers to no one but you.