SMARAN.AI is your personal AI assistant for PC and phone. Wake it by voice and just talk: it understands your screen, opens and controls your apps, plays your music, and builds websites, images and video for you. Run models locally, and your data never leaves your machine.
>
Assistants usually mean a browser tab and a monthly bill. This one is a program on your computer that happens to be very good at talking.
A real-time voice session with interruption handling — start speaking and it stops to listen, the way a conversation actually works. Pick a voice and a persona; it answers in the language you set. The voice that runs on your machine speaks English; other languages are spoken by an online voice, so those need a connection.
Point the camera at a thing, or drop in a screenshot. Vision models read what is in front of them.
Hand it PDFs, notes, spreadsheets. It answers from those, and says so when the answer is not in them rather than inventing one.
Give it a task and watch a real browser window work through it - searching, reading, clicking - with every step shown. It never types passwords and stops before paying or sending anything.
Say "Hey SMARAN", "Hey Amarya", "Hey Myra" or "Hey Jarvis" and it answers "How may I help you?" - play, pause, skip an ad, open an app, look at your screen, draw in Paint. On the phone all four names are heard offline; on the PC "Hey SMARAN" and "Hey Jarvis" are. Your music keeps playing. It never pays, buys or restarts anything without asking you.
Scan a QR code once. After that the conversation follows you between devices, and either one can drive the other — speak to your phone, and your computer does the work.
Enable a startup PIN in Settings, with attempt limits and lockouts to protect access to your history.
CPU, memory, GPU and battery from the actual machine — including whether it is charging — not a decorative animation.
Download local models and keep them forever, or plug in a key for a cloud provider when you want the big ones. The hub is honest about the difference between catalogued, downloaded and actually running.
Give it a goal - "open Notepad and write the list" - and it looks at the screen, clicks and types with your mouse and keyboard, and looks again, every step shown live with its screenshot. Tell it something while it works and it changes course; in Notepad and other plain text boxes it reads the text back to check the typing really landed. Stop ends it at once. While a window that deletes files or changes disks is open, it will not click or type at all. It never types a password or an OTP, and anything about money, buying or the power button comes back to you as a question.
With Web on, SMARAN searches - one quick search for a simple question, several for a comparison or research, the news for what happened today - reads the best pages, and every sentence carries a numbered citation you can open, with follow-up questions to tap.
SMARAN Code reads your project, edits files, runs the tests and fixes what fails. Approvals are Manual, Smart or Off; type an instruction with Allow or Deny; undo a whole run. Start the next task while the last one keeps working.
Make them on this PC and keep everything private, or use the free cloud: pictures from FLUX (a free NVIDIA key) or Pollinations, and short video from Wan 2.2 or LTX-Video on Hugging Face's shared free GPUs - and a finished video can be continued from its last frame. Nothing in it costs money; the free cloud has a daily allowance, and busy hours mean a queue.
The assistant is not a blinking cursor. Open Speak and she is standing there, breathing, watching, answering out loud — rendered live on your own machine, not streamed from anywhere.
Chat, voice, vision and your documents are tabs of the same thing, not four separate apps. Switch mid-conversation and it keeps the thread.
Press once and talk. It hears you while it is still speaking, and stops when you cut in — no push-to-talk, no waiting for a beep.
Filter by what you need — reasoning, vision, code, audio, video. Download a local one and it is yours forever, or point it at a cloud provider with your own key.
Scan the code once. After that, speak to your phone and your computer does the work — or leave the computer off and the phone carries on alone.
Drop in PDFs, notes or spreadsheets. Answers come from those pages — and when the answer is not in them, it says so instead of inventing one.
Memory that carries across sessions, skills it writes down from what worked, a Telegram bot on your own bot token, and scheduled background jobs. Commands run with filters and time limits, and files are checkpointed first so a run can be rolled back. This is not an operating-system sandbox, and the page does not pretend it is.
Run the installer. Python and the app are bundled, so there is no environment to prepare. To run models locally you also need Ollama, which is a separate free install.
Open the Model Hub and download one. A small one runs on almost anything; a bigger one wants more memory. It tells you which is which.
Type, or press Speak. If you want it on your phone too, scan the pairing code.
Every feature included, no card. Local models you download are yours to keep and cost nothing to run.
If Windows 11 blocks the installer with Smart App Control (error 4551), try Download Portable (.zip) above: extract it to any folder and run SMARAN.AI.exe. SMARAN.AI is not code-signed yet, so Smart App Control can block that too; its settings are in Windows Security → App & browser control.
Android will ask you to allow installing from your browser. When download finishes, swipe down your notification shade or open your phone's Downloads app to tap and install.
One file. Make it executable and run it — no package manager,
no install:
chmod +x SMARAN.AI-x86_64.AppImage && ./SMARAN.AI-x86_64.AppImage
Prefer your package manager? Same app, same build: .deb for Debian and Ubuntu, .rpm for Fedora, RHEL and openSUSE, or a .tar.gz to unpack anywhere.
One difference from the Windows build. On Linux the app opens in a browser window with no address bar and no tabs, rather than in a native window. Windows has a webview built into the system; the Linux equivalent, WebKit2GTK, ships in two incompatible versions across distributions and cannot be bundled to work on all of them. Chromium, Chrome, Brave, Edge or Vivaldi will each give you that clean window; with none of them installed it opens in your default browser instead.
smaran ask, chat, models, statusIt is a client, not a second copy of SMARAN.AI. Model routing, your provider keys and the history stay in the app, so the two can never disagree about which model answered.
Step by step, for the platform you are on. Pick one.
A single .exe, about 220 MB. It contains everything the app needs to start.
Windows will stop it, because the app is not code-signed yet. If it says “Windows protected your PC”, that is SmartScreen: choose More info, then Run anyway. If it says “An Application Control policy has blocked this file”, that is Smart App Control, and there is no button to click past it — see the note on the download page before you decide what to do about that.
The installer puts SMARAN.AI in your user folder. No administrator rights are asked for and nothing is written to system directories.
Models run through Ollama, which installs separately. Without it the app still runs and can use cloud providers with your own key.
First launch detects your hardware and suggests models that fit. Nothing downloads until you choose one.
About 77 MB, for Android 7.0 and newer.
Go to downloadsAndroid will ask whether to permit installing apps from your browser. It asks because this is not from the Play Store — that prompt is Android doing its job, not a fault.
The phone app works on its own for voice, chat and camera.
Pairing lets conversations sync both ways and keeps heavy models on the computer, where there is memory to run them. The phone stays usable while the PC does the work.
About 338 MB, for Linux x86-64 with glibc 2.28 or newer (Ubuntu 20.04, Debian 10, RHEL 8, Fedora 29 and later). No package installation is needed. Chats, downloaded models and settings remain in the app data directory.
Go to downloadsA downloaded file has no permission to run until you give it one. That is Linux doing its job.
chmod +x SMARAN.AI-x86_64.AppImage
./SMARAN.AI-x86_64.AppImageThe same build, if you would rather it appeared in your package list:
sudo apt install ./smaran-ai_amd64.deb # Debian, Ubuntu, Mint
sudo dnf install ./smaran-ai.x86_64.rpm # Fedora, RHEL, Rocky, AlmaThere is a .tar.gz too, to unpack anywhere and start with ./run.sh.
The interface opens in a browser window with no address bar and no tabs, rather than in a native window. Windows has a webview built into the system; the Linux equivalent ships in two incompatible versions across distributions and cannot be bundled to work on all of them. Chromium, Chrome, Brave, Edge or Vivaldi each give that clean window; with none of them installed it opens in your default browser.
This release needs glibc 2.28 or newer — Ubuntu 20.04, Debian 10, RHEL, Rocky and
Alma 8, Fedora 29, openSUSE 15.1 and everything released since. Run ldd --version to see yours.
A terminal client for the app already running on your machine. It is not a second copy: the model routing, the provider keys and the history stay in the desktop app, so the two can never disagree about which model answered.
The command line talks to it over 127.0.0.1. Without the app running there is nothing to talk to.
smaran.exeOne file, about 10 MB. Download directly or run in PowerShell:
curl.exe -L -o smaran.exe https://github.com/SHASHWAT-MISHRA-997/SMARAN.AI-downloads/releases/latest/download/smaran.exePut it somewhere on your PATH and the smaran command works.
smaran status # is the app running, and as whom
smaran models # what the hub holds, what is downloaded
smaran ask "..." # one question, answer streamed back
smaran chat # keep asking, in one conversationBy default, requests use loopback and the app treats them as your local device.
An explicit SMARAN_URL can select another backend, whose authentication still applies.
Ask about the file you have open, propose an edit, run a command — without leaving the editor. Connect directly to a local model server or use a cloud provider with your own key.
Free and MIT-licensed.
Open it straight in your editor
Each of these opens your editor at the extension page, where one click installs it.
Rate it, tell us what is missing, or report something that went wrong.
Most assistants say "we care about your privacy" and then send every word you type to a server. Here is what this one actually does.
Yes. Every feature is in the free download. There is no paid tier, no usage cap and no card anywhere in the app. If you connect a cloud provider with your own key, that provider bills you directly — nothing to do with this app.
Downloading the app and local models needs a connection. Once the required models are installed, local chat and supported local speech engines can work offline. Cloud models, Gemini voice and web search need internet access. On Android, dictation depends on the phone’s speech service and installed language packs; that service may require a connection.
It depends on the model, and the hub is specific about it. Across the 63 models in the catalogue the stated memory requirement runs from 4 GB to 512 GB, and 18 of them ask for 8 GB or less. Every entry shows its requirement before you download it.
Yes. The phone app is a working assistant on its own. Pairing adds shared history and lets each device drive the other, but it is not required.
In your own user folder on your own drive. Nothing is uploaded, and no online account holds it. Deleting the app's data folder deletes it.
Built and maintains SMARAN.AI
SMARAN.AI is not a company product. It is built and maintained by one person, on the idea that an assistant should run on your own machine. Questions and bug reports are welcome.