DeskTray Assistant
Chat with an AI model that runs on your own device: ask questions, have it write and explain code, draft emails or summarise text, and keep the conversation going. It also opens the right DeskTray tool, works out calculators in the chat and handles files you attach.
- The optional AI model downloads once from Hugging Face; what you type and attach never leaves this device
- Free, no sign-up
- Works on phones
You’re offline. This tool needs an internet connection for some features. They will work again when you reconnect.
AI is off: get a model in Settings to chat, write and answer questions
On-device models are much smaller than online chatbots: they make more mistakes and can state wrong facts with confidence, so check anything important. They can’t browse the internet. Calculations use the site’s own calculators, and I show the numbers I used.
Hi! I’m the DeskTray Assistant, an AI chatbot that runs on your device. Ask me anything, have me write code or an email, or tell me what you want to do on the site: I’ll open the right tool, work out a calculation or handle a file you attach. Nothing you type or attach is uploaded.
To chat, get an AI model in Settings (the AI bar above links there); it downloads once. Without it I find tools and work out calculators.
What goes online
- The optional AI model downloads once from Hugging Face; what you type and attach never leaves this device.
- Everything else happens inside this browser tab.
Often used next
Looking for something else?
DeskTray has hundreds of free tools that work the same way.
How to use the DeskTray Assistant
- Turn on the AI once: download a model in Settings (the AI bar above the chat takes you there; the recommended one suits this device). It is kept in your browser and then works offline.
- Type a message, such as “Write a Python program to find prime numbers” or “Explain how compound interest works”, and press Enter. The reply streams in; Stop ends it early.
- Keep chatting: it remembers the conversation. Continue finishes an answer that was cut off, Regenerate writes the last one again, and Copy works on any reply or code block.
- Attach a file and ask about it (“explain this mail”), ask for a tool (“merge PDF files”) or give a calculation (“EMI for 20 lakh at 8.5% for 20 years”): it reads and answers, opens the tool or works it out in the chat.
Features
- A chatbot that runs on your device: answers questions, writes and explains code, drafts emails and summarises or translates text
- Remembers the conversation, with a choice of memory from 4K to 32K tokens, and summarises older messages when it fills
- Formatted replies with code blocks you can copy, plus Stop, Continue, Regenerate and an optional “think before answering” mode
- Saved chats in your browser, with custom instructions and a choice of precise, balanced or creative replies
- Opens the right DeskTray tool when you ask for one, and works out 40 calculators such as EMI, GST, SIP and income tax in the chat
- Reads files you attach and answers about them: PDFs (scans too), Word, spreadsheets, code and the words in pictures
- Compresses photos to a KB limit, converts images, makes a PDF from photos, merges PDFs and opens attached files in the right tool
A chatbot that runs on your device
Turn on the AI and the assistant becomes a chatbot. It answers your questions and does the writing itself, so “write a Python program to reverse a string” gets you the program, not a list of tools. The model runs in your browser on your device’s graphics chip, so it works offline once downloaded and nothing you type is sent anywhere.
- Code: programs and functions in Python, JavaScript, Java, C, SQL and more, explained step by step, and fixes for code you paste in.
- Questions and explanations: how something works, the difference between two ideas, a topic explained simply for a student.
- Writing: emails, leave applications, letters, messages and posts, and summaries, rewrites, corrections or translations of text you paste.
- Follow-ups: “make it shorter”, “add comments to the code”, “now in Hindi”. The assistant remembers the conversation, so you don’t have to repeat yourself.
How much it remembers: the context window
A language model reads the conversation each time it answers, and it can hold only so much at once: its context window, counted in tokens, which are pieces of words (a token is about three quarters of an English word). The window holds the instructions, the conversation and the answer being written.
You choose the window in Settings, under the text models’ Memory: 4K tokens (about 3,000 words), 8K, 16K or 32K (about 24,000 words). Automatic picks 8K on computers with memory to spare and 4K on phones and smaller devices. A larger window remembers longer chats, but the model reserves graphics memory for it as soon as it starts, about 112 MB for every extra thousand tokens on Qwen3 1.7B, so the list shows what each size needs and greys out sizes this device can’t hold.
The bar under the message box shows how full the memory is. When a chat no longer fits, the oldest messages are summarised in a few lines that stay in the model’s memory, the chat notes it, and the conversation carries on. Details from those early messages can be lost, so for a new topic it is better to start a new chat.
Replies, code and controls
Replies are formatted: headings, lists, tables and code blocks with the language shown and a Copy button that copies just the code. Stop ends an answer early, Regenerate writes the last answer again, and Copy under an answer copies all of it. A long answer can stop at the model’s limit; Continue then picks up where it stopped, inside the same code block if it was writing code.
In the chat settings, “Think before answering” lets the Qwen3 models reason before they reply. It helps with maths, logic and trickier code, takes longer, and the reasoning can be opened above the answer. Replies can be precise, balanced or creative, which sets how freely the model chooses its words, and custom instructions are added to every chat, such as “I’m a Class 10 student; explain simply with examples in rupees”.
Tools, calculators and files
When you clearly ask for a tool, such as “merge PDF files”, “EMI calculator” or “PNG to JPG”, the assistant shows the matching DeskTray tools with an Open button, set to the right mode where a tool has several. When you ask a question about a job, such as “how do I merge PDF files?”, it answers and links the tool under the answer.
Forty of the site’s calculators run inside the chat: loans and EMI, GST, SIP, lumpsum, FD, RD, PPF, NPS, income tax, in-hand salary, HRA, gratuity, percentage, BMI, age and more. The assistant reads the numbers from your message, including Indian ways of writing them such as 20 lakh, 1.5 crore or 50k, and runs the same calculation as the calculator’s own page, because a small language model is not reliable at arithmetic. Every result lists the values used, and a reply with just a new number, such as “9.2%”, recalculates.
Attach a file with the clip button and ask about it, such as “explain this mail”, “what is the last date?” or “summarise this report”, and the assistant reads it and answers. It reads PDFs (scanned pages with OCR), Word documents, spreadsheets, text and code files, and the words in pictures; the on-device model reads text, so it can’t describe what a photo shows. A long file is answered from the pages that match your question, a summary of one is read in parts, and later questions in the same chat are answered from the same file. Under the answer, the tools for that file are ready to open with it.
Ask for a job on a file and it is done instead: a photo under a size in KB, images converted to JPG, PNG or WebP, photos put into one PDF and PDFs merged happen right in the chat, with the same code the tools use. For everything else, such as compressing, splitting or signing a PDF, “Open with file” starts the right tool with your file already loaded.
Saved chats
Chats are saved in this browser, so you can come back to them: Chats lists them with the newest first, and each can be opened again or deleted, or all deleted at once. They are stored in the browser’s own database on this device, never uploaded, and they go if you clear the site’s data. You can turn saving off in the chat settings; the open chat then lasts until you close the tab.
With and without the AI model
The assistant works straight away, without any download, as a tool finder: it matches your words against every tool’s name, description and keywords, including everyday words such as “pic”, “EMI” or an exam name, runs calculators whose numbers are clearly marked, and does the file jobs above. Chatting, writing and code need the model.
The models are Qwen3 0.6B, 1.7B and 4B, Google’s Gemma 3 1B, Gemma 4 E2B and Gemma 2 9B, and Ministral 3 3B, from 337 MB to 4.9 GB, downloaded once from Hugging Face. The recommended one depends on your device. Larger models answer better and know more but write more slowly; Qwen3 is also the better choice for Indian languages. They need a browser with WebGPU.
Private by design
Online chatbots send what you type to a company’s servers. This one doesn’t: the only things it downloads are the list of DeskTray tools from this site and, if you choose, the AI model from Hugging Face, with its small graphics-chip library from GitHub. Your messages, numbers and files stay on your device, and the model can’t send them anywhere.
What a small on-device model can’t do
Models small enough for a phone or laptop are far smaller than online chatbots. They know less, make more mistakes in facts, maths and code, and sometimes state something wrong with confidence, especially in Indian languages on the smaller models. Their knowledge stops at their training date, and they can’t browse the internet, open links or see the news. Treat answers as a draft: test code before you use it, and check anything important against a reliable source.
Frequently asked questions
Is what I type or attach uploaded?
No. The assistant and its AI model run in your browser. It loads the list of DeskTray tools from this site, and the model, when you turn it on, downloads once from Hugging Face. Your messages and files are processed on your device, and saved chats are kept in this browser only. When it opens a tool with your file, the file is handed over through your browser’s own storage, where it is kept for a day at most.
Do I need to download the AI model?
For chatting, yes: answers, code and writing come from the model (337 MB to 2.1 GB, downloaded once). Without it the assistant still finds and opens tools from your words, works out calculations whose numbers are clearly marked (₹ or lakh for amounts, % for rates, years for periods), and does file jobs such as compressing a photo to a size in KB or merging PDFs.
How much of the conversation does it remember?
As much as fits in the model’s memory, its context window: 4K tokens (about 3,000 words) to 32K tokens (about 24,000 words), shared by the conversation and the answer being written. It starts at 8K on computers with memory to spare and 4K on phones, and you can change it in the AI bar. When a chat fills the memory, the oldest messages are summarised in a few lines, so the conversation can go on.
Can I trust its answers and code?
Check them. Models small enough to run on a phone or laptop make more mistakes than large online chatbots, and they can state wrong facts confidently or write code with bugs, so test code before you rely on it. Calculations are different: they come from the same calculators as the rest of the site, and the assistant lists every value it used.
Which browsers can run the AI?
The AI models need WebGPU: Chrome or Edge on Windows, Mac and Chromebooks, Chrome on recent Android phones, and Safari on iPhone, iPad and Mac with iOS or macOS 26 or later. In other browsers the assistant works without the AI.
Why did it show tools instead of answering?
A short request that names a job a tool does, such as “merge pdf files” or “passport size photo”, opens the tool, because the tool does the job better than a chat reply. To get an answer instead, ask a question (“how do I merge PDF files?”) or say what to write (“write…”, “explain…”). Without the AI turned on, the assistant can only find tools.
Can it read the files I attach?
Yes, with the AI turned on. Ask a question with the file, such as “explain this mail”, and it reads the text of PDFs, Word documents, spreadsheets, text and code files, and the words in pictures and scanned pages (with OCR), then answers. Everything is read on your device; the file is not uploaded. Without the AI it offers the tools that work with the file.
Why were earlier messages summarised?
The chat no longer fitted in the model’s memory (its context window), so the oldest messages were condensed into a short summary to make room. Choose more memory in the AI bar for longer chats, or start a new chat for a new topic.
Where is the assistant on other pages?
The round button at the bottom right of most pages opens it in a panel. It is hidden on games and full-screen tools such as the video editor, where it would get in the way. The chat is the same in the panel and on this page.