Local-first AI for India: your work stays on your computer

If you run a CA practice, a small export business or a freelance studio in India, your computer holds some of the most sensitive data in your life: PAN cards, Aadhaar copies, GST returns, bank statements, client contracts. An AI assistant that wants to help with that work should not need to carry it off somewhere else first.

The short answer: a local-first AI agent keeps your chats and memory on your own computer by default, and leaves your files where they are, so your work stays wherever you and your computer are. Pivra Desk works that way, can hide identifiers like Aadhaar, PAN and GSTIN before anything reaches a cloud model you choose, and keeps working on a model on your computer when the internet doesn't.

This is the second post in our series for India. The first one sets out the wider stand.

What "local-first" means for your data

Pivra Desk keeps your agents, chats, memory, rules and routines in a local database on your computer. Your logins and keys are sealed in a vault on the same machine. Nothing about your work is sent to Pivra by default, and we never train on your data.

That has a simple consequence. If your laptop is in Pune, your work is in Pune. It does not sit in a server farm in another country waiting to be processed.

To be precise about what can leave, and when:

• A model on your computer: your conversation stays on the machine. Nothing goes to a model provider.
• A cloud model you chose: only when the agent uses it, from your computer straight to that provider under your own account. That provider's servers may be outside India; that is their terms, not ours.
• Web search and reading a page: the search words go to DuckDuckGo; a page's address goes to that site.
• Connected apps such as Google: only when the agent uses them, straight to that service.
• Also: update checks, and your plan check if you sign in, but never your work.

The full list is on Privacy, and where your data lives and the security page. We would rather you read it than take our word for it.

Hiding Aadhaar, PAN and GSTIN from cloud models

Sometimes a cloud model is the right tool, for example to draft a long notice reply or to compare two contracts clause by clause. But the document may contain a client's Aadhaar number or a supplier's GSTIN.

Pivra Desk has an option for exactly this. In Settings → Privacy, under When a cloud model is used, choose Hide identifiers. Before anything goes to the cloud model, email addresses, phone numbers, card and bank numbers, and Aadhaar, PAN, GSTIN and UPI IDs are replaced with labels like «EMAIL_1». The real values stay on your computer and are put back into the reply, and before anything runs.

Four things to know:

1. It is off unless you turn it on. It is your choice, per computer.
2. It hides identifiers; it does not make a chat anonymous. The cloud model still sees the rest of your text: names, amounts, the story of the case.
3. It only matters for cloud models. With a model on your computer, your text doesn't go to a model provider in the first place.
4. It works on text, not pictures. A scanned card or a photo is sent as it is. Identifiers are found by their format and check digits, so read before sending anything unusual.

For a Chennai CA practice, a sensible pattern is: routine reading and summaries on the local model, and Hide identifiers switched on for the occasional cloud job.

Rupees, lakh and crore, done exactly

Indian money is written in lakh and crore, where the last three digits are grouped, then every two (12,34,567). A language model, local or cloud, is not a calculator. It can get "18% GST on ₹12,500" subtly wrong, and in invoices "subtly wrong" is just wrong.

Coming in the next update: a question that is just a sum is worked out exactly, in code, on your computer, with no model involved. For sums inside a longer job, the model gets a calculator, and a number in the reply that the calculator didn't produce is replaced by the computed one. That covers:

• amounts in ₹ (and Rs or INR);
• lakh and crore, and Indian grouping such as 12,34,567;
• percentages and GST, for example "18% GST on ₹12,500" gives the tax (₹2,250) and the total (₹14,750);
• Hinglish phrasing such as "2340 ka 15% kitna hai".

A Kolkata trading firm checking a supplier's invoice total, or a Kochi freelancer working out TDS at a rate they give, should be able to trust the arithmetic. That is what this change is for.

Hindi and Hinglish

Many people switch between English, Hindi and Hinglish in the same sentence. Today you can write to Pivra Desk in Hindi or Hinglish, and whatever you write stays on your computer when you use a model there.

Coming in the next update: Hindi and Hinglish requests get the same tools as English ones (reading files, searching, drafting), and your agent's search of your chats and memory gets more reliable in Hindi. We have not published measurements of how well each model writes Hindi, so we make no claims about that yet. The last post in this series goes into detail.

Works on patchy internet

Connections drop. Mobile data runs out. The broadband in a new office takes three weeks to arrive. A local-first agent copes with that.

• The model on your computer needs no internet once it's downloaded. Reading, summarising and drafting carry on offline.
• Download it once, on good Wi-Fi. Gemma 4 E2B, the model we suggest for 8 GB computers, is a 3.3 GB download; the app itself is about 170 MB. After that, the model is on your disk.
• Your plan keeps working offline. If you're on Plus, the app checks your plan on your computer and keeps it on through a grace period while it can't reach pivra.ai.
• What needs a connection: web search, connected apps and cloud models.

Asks before it acts, in every language

None of this changes the rule underneath. Reading and searching run on their own. Sending, publishing, paying and changing anything wait for your approval on a card that shows exactly what will happen. Money, deletes and deploys need Touch ID on a Mac or a phrase you type. You can add rules in plain words. How approvals work.

Frequently asked questions

What is Pivra Desk?

Pivra Desk is a local-first personal AI agent for Mac and Windows. It works across your files, browser and apps using a model on your computer or a cloud model you choose, and asks before consequential actions.

Does my data leave India?

With a model on your computer, your chats and memory stay on your computer and your files stay where they are; if the agent searches the web, only the search words go to DuckDuckGo. If you choose a cloud model, your requests go to that provider, whose servers may be outside India. You decide which to use.

Can I stop Aadhaar or PAN numbers reaching a cloud model?

Yes, in text. Turn on Hide identifiers under Settings → Privacy → When a cloud model is used. Identifiers are replaced with labels before the cloud call and restored on your computer. It does not hide the rest of your text, and it does not read pictures such as a scanned card.

Does it understand lakh and crore?

Coming in the next update: a question that is just a sum in ₹, lakh or crore is worked out exactly, in code, on your computer.

Is there a price in rupees?

Not yet. There is a free plan, and Plus is US$6 a month, charged in US dollars outside Australia; taxes may apply. See pricing and plans.

Which computers does it run on?

Macs with Apple silicon and 64-bit Windows 10 or 11. Install guide.

Try it on real work

Download Pivra Desk, point it at a folder of last quarter's invoices, and ask for a summary. Your files stay on your computer, and nothing is emailed, published or paid without your yes.