AI

AI that drafts, and a store that answers questions.

Two separate capabilities, and it is worth keeping them separate. A connector lets Claude or ChatGPT read your real data and answer in plain language — on the subscription you already pay for, with no API key and nothing billed per token. A key is the second path, for the work that has to run when nobody is watching. Neither can publish anything you have not approved.

In the admin

AI writes the copy you keep putting off

Product descriptions, SEO meta, alt text, FAQ blocks, broadcast and journey messages. It reads your existing published copy first, so drafts sound like your brand rather than like a model that has never seen your catalog.

Everything lands in a review queue. You apply it or you reject it.

Outside the admin

Ask Claude about your own store

Connect Claude or ChatGPT once, then ask "which products are quietly running out of stock?" or "what did the Diwali campaign actually earn?" and get an answer from your live database — not from a dashboard you have to interpret yourself.

50 tools, no API key, nothing billed per token. Read-only until you decide otherwise.

Which one

The connector is the cheap one. Here is why.

You almost certainly already pay for Claude or ChatGPT. The connector uses that subscription — the model runs on your side and the store just answers its questions, so nothing is billed per token at all. The API key is a separate path, and you only need it for the jobs that have to run when nobody is watching.

  Connector — your Claude / ChatGPT subscription AI API key — billed per token
What pays for it The subscription you already have. No new account. A key from Anthropic, OpenAI, Gemini or any compatible provider.
Cost per task ₹0 — the store spends no tokens. It only hands over data and stores what you wrote. Fractions of a rupee, recorded per task and capped daily.
Where the model runs In Claude or ChatGPT, on your side. Called from your store’s own server.
A key stored in the store None needed Required, kept in settings.
Ask questions of live data Yes — this is the whole point of it. No. The key powers generation, not a chat window.
Write product copy, meta, alt text Yes, into the same review queue. Yes, into the same review queue.
Runs unattended No — you are in the conversation. Yes — queued jobs, no one present.
Answers customers on WhatsApp No Yes — the concierge needs this path.
Bulk work across the catalog Product by product, in the chat. Queued and processed in the background.
Setup Connect once. Claude and ChatGPT register themselves. Paste a key, pick a model, set a daily cap.
Start with the connector. For writing copy, auditing SEO and asking where the money went, it costs you nothing beyond the subscription you already pay for. Most stores never need the API key at all.
Add the key when you want it working while you sleep. The WhatsApp concierge and any bulk or scheduled generation have to run on the server, and that is the one thing a chat session cannot do.

Both are included in the ₹30,000 build. Use either, both, or neither.

The part that matters

Three tiers, and you control where the line sits.

Handing an AI model access to a shop is only sensible if the blast radius is small and obvious. It is enforced in code, not in a system prompt a clever message can talk around.

Read

35 tools that only ever look. Each one checks the same staff permission as the admin screen showing the same data — so the connector can never reach something the person who connected it could not already open.

Draft

8 tools that write to the review queue and nowhere else. A human still has to apply the draft before a single character reaches the storefront. These stay available even when writes are switched off, because they cannot change anything live.

Write

7 tools that change live data — stock, coupon state, review moderation, publishing. Off by default, behind a separate switch from the connector itself, and every one of them leaves an audit trail marked as having come from the connector.

Why the tiers exist. A connector token reaches orders, customer contact details and revenue. If a hostile instruction ever reaches the model — hidden in a review, a product enquiry, a page it was asked to read — the worst it can do is disclose what the token's owner could already see. It cannot place an order, refund a payment, change a price, export your customer list or grant itself anything. Those tools do not exist.
Concierge

It answers WhatsApp at 11pm so you don't have to.

Customers ask the same eight questions forever. The concierge answers them from your live catalog, on WhatsApp and Instagram — DMs and comment replies both.

  • It cannot invent a product, a price or a link. Your catalog is searched first and the model may only pick from that shortlist — any id it returns that was not on the list is dropped. The product card is then written by the store from the catalog row, not by the model.
  • It knows their last order. The order number and its status, resolved through the identity graph and only from a verified contact. A WhatsApp number is verified by Meta, so their orders surface; an Instagram handle with no verified contact returns nothing rather than guessing at a stranger's order.
  • It hands over cleanly. Both sides of every conversation are logged into the same inbox your team already uses, so a person can take the thread at any point.
  • It degrades safely. With AI switched off, the key removed, or the provider simply having a bad day, it falls back to a helpful fixed reply instead of failing in front of a customer.
WhatsApp 11:04 pm
hi! do you have anything in linen under 3k?
We do — this one’s a favourite and it’s comfortably under your budget. Want me to show a couple more in linen? Sandwash Linen Shirt — ₹2,450
yourstore.in/p/sandwash-linen-shirt

Only the first two lines are the model’s. The name, price and link below the divider are appended by the store itself, straight from the catalog row — so they cannot be wrong even if the model would like them to be.

Cost

If you do add a key: your account, at cost.

The connector needs none of this. But when you want the concierge answering at 3am, or generation running across the whole catalog, you bring a key from whoever you like. We do not resell tokens and we do not mark them up.

₹0 Added by us, on either path. No per-word plan, no AI tier, no resold tokens.
Daily cap A rupee limit checked before every call. A loop cannot run up a bill overnight.
Per task Tokens in, tokens out and the rupee cost recorded against every single generation.
Anthropic OpenAI Google Gemini Groq Together DeepSeek Mistral OpenRouter Ollama, on your own machine

Anything speaking the OpenAI-compatible API works with a base URL, a key and a model id. Running a local model means your catalog never leaves your server at all.

The connector

What Claude can actually see.

Every tool, grouped. Reading tools obey your staff roles; drafting tools write only to the review queue; the writing tools stay off until you switch them on.

Connecting is one paste of a URL: Claude and ChatGPT register themselves over the remote MCP standard, using OAuth 2.1 with PKCE. Tokens and client secrets are stored hashed, and a token issued for another service cannot be replayed against yours.

Sales & revenue

8 read-only tools

Headline KPIs for any date range
Revenue per day, gaps zero-filled
Best sellers by units and revenue
Revenue split by acquisition channel
Conversion funnel, session to order
Coupon redemptions and discount given
Ad spend against attributed revenue
Traffic by source, with a live snapshot

Orders & customers

6 read-only tools

Order counts by status
One order in full, by order number
Customers ranked by lifetime spend
Abandoned carts with contact details
How each recovery message is performing
Everything about one customer, across channels

Catalog & stock

4 read-only tools

Search the catalog by name
Variants at or below a stock threshold
Units on hand valued at retail
Price bands against where sales actually happen

Marketing & demand

3 read-only tools

Broadcast delivery, queued through read
Journey runs: active, completed, exited
What shoppers typed into your search box

Shipping

1 read-only tool

Courier status, RTO rate and stuck parcels

Reviews

1 read-only tool

Reviews awaiting moderation, oldest first

SEO & content research

12 read-only tools

Measured SEO coverage across the catalog
Your brand voice, from published copy
What copy is missing, by area
Everything needed to write one product
Copy from products that already sell
Demand the current copy does not mention
How a title and meta will look in Google
What structured data each page emits
Pages and products nothing links to
Images with missing or useless alt text
A page's sections and their current copy
The actual product photographs, to look at

Drafting

8 tools. These write to the review queue only — a human applies them.

Product copy
Category copy
Page title and meta
Page-builder section copy
Broadcast message
Automation step message
Image alt text
FAQ block

Writing

7 tools that change live data. Off by default, audited when on.

Approve or reject a review
Activate or deactivate a coupon
Adjust stock on hand
Publish or unpublish a product
Publish product copy
Publish category copy
Publish page title and meta

And if you want none of it?

Configure no key and every one of these features sits dormant. The connector stays off, the concierge falls back to a fixed reply, and the store runs exactly as it otherwise would. None of this is a subscription, and none of it is switched on behind your back — it is all included in the ₹30,000 build either way.

Next step

Look around it before you spend a rupee.

A fixed price, agreed before anything begins, and the source code in your hands at the end of it. If it isn't right for you, we'll say so.

One-time payment Your hosting, your choice Source code handed over No revenue share