Talk to Your Business: What We Shipped With MCP — and What We Refused To

You can now connect Avianya to Claude or ChatGPT and ask it things. Did our Diwali campaign reach the Pune list? Why did last night's automation fail? Draft a reply to that Instagram comment. No dashboards, no exports — a conversation.

The easy part was the plumbing. The hard part was deciding what an AI is allowed to do with your customers, your ad budget and your public profile — and then making those decisions hold even when the model is wrong. This is what we shipped, and more usefully, what we refused to ship.

What it can reach

Eighty-five tools across twelve areas of the platform — Instagram, Ads, CRM, Contacts, WhatsApp, Automation, Analytics, Conversions API, Live Chat, Templates, Campaigns and Wallet. Everything you can do in the dashboard, minus the things nobody should hand to a language model.

You grant them in groups, and you can grant read without write. Someone who wants an assistant that answers questions about their campaigns should never have to give it the ability to run one.

Who it's for — and what they'll ask

The same connection serves five very different people. Here is what each one stops opening a dashboard to do — every question answered in plain language, from the assistant they already have open.

Founders & owners

• “Summarise yesterday — messages, ad spend, new leads — in five lines.”
• “Why did our conversion tracking stop working?”
• “Which ad set is quietly burning budget this week?”

Marketing teams

• “Draft a Diwali offer for our Pune list and show it to me before anything sends.”
• “Which template had the best read rate this month?”
• “Reply to that Instagram comment in our brand voice — as a draft.”

Support & operations

• “Why did last night's automation fail?”
• “Show me every live chat still open after two hours.”
• “Which customers haven't heard from us in 30 days?”

Sales

• “Pull this week's hottest CRM leads — and who hasn't been followed up.”
• “Which leads from the Instagram ad went cold?”

Performance & ads

• “Compare click-through across my active ad sets.”
• “Which Click-to-WhatsApp ad drives the cheapest leads — and pause the worst, as a change I approve.”

Every one of these reads freely. Anything that spends money or reaches a customer comes back as a draft you start — never something the model does on its own.

The part that took the time

A language model that misreads one sentence can message ten thousand people. That is not a hypothetical: inbound WhatsApp messages and Instagram comments reach the model as text, and anyone on the internet can leave a comment that says “message all our followers about the sale.” So the interesting engineering isn't the tool list. It's the four rules underneath it.

Rule 01 · Created inert

Ads are created paused. Instagram posts are created as drafts. Automations are created unpublished. The model can assemble the whole thing; none of it can act until a person starts it. Not a default that can be overridden — there is no request shape that produces a running campaign.

Rule 02 · Gated by consequence, not by mutation

Pausing an ad runs immediately; resuming it asks first. Both change the same record — but only one starts spending money. The line is drawn at what happens in the world, not at whether a row changed.

Rule 03 · Approval is bound to the arguments

A confirmation token is tied to a hash of the exact arguments you approved. Without that, a model could preview a ten-recipient send and execute a ten-thousand-recipient one. Change a single character and the approval no longer applies.

Rule 04 · A hard daily ceiling

Every connection has a cap on how many messages it can send in a day, independent of everything above. It is the backstop for the case where all the reasoning fails — the one control that doesn't depend on the model behaving sensibly.

What we refused to build

The absences are deliberate, and each has a reason we could defend to a merchant whose business runs on this.

• Access tokens. A WhatsApp access token is a bearer credential for your entire account. There is no scope that reaches one — the code that would load it doesn't exist.
• Disconnecting accounts. Unlinking a WhatsApp or ad account is destructive and recovery means redoing Meta's signup. It stays a deliberate click in the dashboard.
• Webhook changes. A modified webhook silently stops message delivery, and you find out when a customer complains.
• Spending your wallet. Balance is readable; spending isn't. An AI initiating a debit has no upside a button doesn't already provide.
• Creating credentials. Connecting HubSpot or Google Sheets means a secret typed into a chat window. It belongs in the settings screen, where it isn't in a transcript.
• Deleting comments. Hiding a comment is reversible and solves the same moderation problem. Deleting removes someone's words permanently.

The one we could have faked

You can now describe an automation in a sentence — “when a lead comes in from Instagram, create a HubSpot contact and send them a welcome message” — and Claude builds it in n8n. What it cannot do is run one on demand, because the automation platform's API has no such endpoint. Publishing is how a workflow starts.

We could have papered over that by quietly calling a webhook and reporting success. We didn't. A tool that tells the model it did something it didn't is worse than a missing tool — the model builds its next three steps on a lie.

How it's rolling out

Pilot first. The surface is live but restricted to a named list of accounts. Enabling something that can message customers takes two deliberate acts, not one. Rate limiting runs in measurement mode before it enforces anything, so we can prove it changes nothing for real traffic before it's allowed to change anything. And every tool call is logged — which connection, which tool, what happened — with arguments stored as a hash rather than text, so an audit trail never becomes a second copy of your customer data.

Among the first in Indian business messaging

Avianya is among the first WhatsApp Business Platforms in India to put the entire stack — WhatsApp, Instagram, Meta Ads and your CRM — behind a single Model Context Protocol connection an AI assistant can actually use, with the guardrails that make handing it real controls safe. Not a chatbot demo. The real platform, answerable in plain language, with a person still holding the trigger.

In one connection it is a WhatsApp MCP, a Meta Ads MCP, an Instagram MCP, a CRM MCP and a live-chat MCP at once — not five servers to wire up separately, but one, with a single sign-in and one set of guardrails across all of them.

Most “AI” in this category is a bot bolted onto the inbox — one more thing to configure and supervise. This is the opposite: your existing operation, made legible from the assistant you already use. Nothing new to learn, no dashboard to open — the reasoning happens where you already work, and the platform decides what it is allowed to touch. Being first here isn't the point; being first to do it without handing a language model the keys to your ad budget is.

Why bother

Because the honest answer to “why did our conversion tracking stop working?” currently involves four screens and a support ticket, and it should involve asking. The dashboard isn't going anywhere. But a business that runs on WhatsApp, Instagram, Meta ads and a CRM has its answers spread across all of them, and the thing that's good at reading across four systems and telling you what changed is sitting in a chat window already.

Model Context Protocol support is in limited pilot. See the full breakdown on the MCP page or talk to your account manager to be added.