You can now ask Claude, "who messaged my business on Instagram this week and never got a reply?" or "why is that competitor growing faster than me?" and get a real answer, pulled from your own account. This is new, most of the ways people are doing it break Meta's rules, and the difference matters more than the setup does.

The short version: AI assistants learned to use outside tools in 2025, the standard for that is called MCP, and Instagram is now reachable through it. What follows is what actually works in September 2026, what will get your account restricted, and how to tell the two apart before you hand an AI the keys to your inbox.

Three Different Questions Hide Behind This One

"Connect Claude to Instagram" means at least three separate things, and the answers are not the same:

  • Read my DMs. You want to ask questions about conversations that already happened. Who asked about pricing, who never got a reply, what people keep asking. This is the safe one and the useful one.
  • Reply for me, live, all day. You want an AI answering customers at 2am without you. This is a different product entirely, and an assistant on your laptop is the wrong shape for it.
  • Post for me. Scheduling and publishing. Possible, and mostly unrelated to the other two.
  • Tell me how I am doing. Your reach, your audience, and how you stack up against the accounts you compete with. The least obvious one, and the one people end up using most.

Most people asking the question want the first one and discover they also want the second. Keep them separate in your head, because the tools that do them are different.

What MCP Actually Is

The Model Context Protocol is an open standard, introduced by Anthropic in late 2024 and since adopted by OpenAI, Google and most of the coding tools. It is a common language for "here are some tools, an AI may call them."

Before MCP, connecting an AI to your data meant someone building a custom integration for that one AI. Now a service publishes an MCP server once, and Claude, ChatGPT, Cursor, Gemini and anything else that speaks the protocol can use it. It is the same idea as a USB port: one shape, many devices.

An MCP server for Instagram, then, is a program that offers an AI a menu of things it may do with your Instagram account. Read conversations. Look up a contact. Send a message. The AI decides which to call and when, and you decide whether to let it.

Two ways an MCP server reaches you

Hosted (remote). You paste a web address into your assistant, click Connect, and sign in. Nothing is installed. This is the only kind that works in claude.ai on the web, in ChatGPT, or on a phone.

Local. A small program runs on your own computer. Works with Claude Desktop, Cursor, Zed and similar. Nothing to sign into, but nothing on your phone either.

The Part That Will Get Your Account Banned

Search for an Instagram MCP server and you will find a pile of free ones on GitHub. Most of them work by logging into Instagram as you, using your username and password, and reading the web interface the way a browser does. Some go through unofficial mobile endpoints.

This is scraping, and it violates Meta's Platform Terms whichever wrapper it comes in. Meta detects automated web sessions well, and the consequences land on the account, not the tool: action blocks, then a restriction, then in the worst case removal. We wrote about the 2026 ban wave after watching it happen to people who thought they were using a normal automation tool.

The tell is simple, and you should check it before installing anything:

What it asks you for What that means Verdict
Your Instagram username and password It is logging in as you and scraping Against Meta's terms. Real ban risk.
A session cookie or "sessionid" Same thing, one step further hidden Against Meta's terms. Real ban risk.
A "connect with Facebook" or "connect with Instagram" button Official Meta login, permissions you can see and revoke Allowed. This is the supported path.

The free ones are free because they skip the part that costs money and time: becoming an approved Meta app, passing app review, and staying inside the API's limits. There is no version of this where a GitHub project you found this morning has quietly solved that.

What Meta Actually Permits

Through the official Instagram Messaging API, an approved application can read and send messages for a business or creator account that has connected itself. That is genuinely permitted, and it is how every legitimate tool in this space works, including ours. We are a Meta Verified Tech Provider, which is the status Meta grants for exactly this.

The rules that come with it are worth knowing, because they explain limits you will otherwise blame on the AI:

  • The 24-hour window. After a customer messages you, you have 24 hours of free-form replies. Outside it, Instagram and Messenger only allow specific message types. This is Meta's rule, not a product limitation. We covered it in the 24-hour window guide.
  • Business and creator accounts only. Personal accounts cannot connect. Neither can a business account that is not linked to a Facebook Page.
  • You granted it, you can revoke it. The connection appears in your Meta settings and dies the moment you remove it.

How the Supported Version Works

The path that stays inside the rules has one more step than the scraper version, and the step is the point. Instead of the AI holding your Instagram password, it holds a token for a service that Meta has already approved, and that service talks to Instagram.

For ChatGenius that looks like this. Instagram is connected to ChatGenius once, the normal way, through Meta's own login. Then you point your assistant at our MCP server, approve it, and the assistant works through us. Your Instagram credentials never touch the AI, and you can cut off either connection independently.

The address is https://sumgenius.ai/api/mcp and it is a hosted server, so it works in claude.ai on the web and on your phone, not only in a desktop app.

What you approve, and how much

When you connect, ChatGenius asks you to choose. Read only lets the assistant see conversations, contacts, appointments, flows and posts, and nothing else. Read and act also lets it send messages, book appointments, build automations and publish. You can pick the smaller one even when the assistant asks for the larger one, and you can disconnect any assistant from your settings at any time.

What You Can Actually Ask It

This is the part that surprises people. Once it is connected, you are not clicking through a dashboard, you are asking questions in plain English and getting work done. Real examples:

  • "Who asked about pricing in the last two weeks and never got a reply?"
  • "Tag everyone who mentioned the September promo as a warm lead."
  • "What time do most of my DMs actually arrive?"
  • "Book Maria for Thursday at 2pm and tell her."
  • "What are people asking that my AI could not answer? Draft FAQ entries for the top five."
  • "Which of my automations gets people to the end, and which loses them?"
  • "Build a flow that answers the word price with my packages and collects an email."
  • "How do I compare to the three accounts I track? Where am I actually winning?"
  • "They post six times a week and I post twice. Is that why they are growing faster?"
  • "What hour is my audience most online, and does that match when I post?"

That last one is worth pausing on. The assistant reads the catalogue of available steps, writes the automation as a structure, checks it for errors before anything is saved, and creates it. You then open it in the visual builder and edit it by hand like anything else. Describing an automation and getting a working draft is a different experience from dragging boxes around.

The Part Nobody Else Does: Competitor Questions

Reading your DMs is the obvious use. The one that surprised us is competitive research, because it is the exact shape of work an AI is good at and a dashboard is bad at.

Instagram will tell you your own numbers. Working out what they mean next to somebody else's takes a spreadsheet and an afternoon. So people do it once, feel bad, and never do it again. An assistant does it in a sentence.

Track a handful of accounts you compete with and the assistant can reach their follower counts, engagement rates, posting cadence and content mix, alongside your own, all from Instagram's official API. Then you ask it things you would never build a report for:

  • "Rank me against the accounts I track. Where am I ahead and where am I behind?"
  • "Their engagement is five times mine on a tenth of the followers. What are they doing that I am not?"
  • "Compare our content mix. They are 60 percent reels and I am 60 percent carousels. Which is working?"
  • "Show me their follower growth over 90 days next to mine and tell me when the gap opened up."

The last one is worth dwelling on. The assistant is not being asked to look something up. It is being asked to notice a change and explain it, which is a genuinely different thing from a chart.

It reaches your own audience data too, the parts Instagram keeps behind a tab most people never open: age and gender breakdown, top countries and cities, and what hours your followers are actually online. "When should I post?" stops being a guess from a blog article and becomes a question about your own audience.

What this is not

None of this is scraped. It is Instagram's public business data through the official API, the same numbers Instagram shows you, and it only covers public business and creator accounts. Nobody can read a private account's analytics, and any tool claiming otherwise is either lying or logging in as you. Competitor slots are capped by plan, from one on the free tier to ten on Business.

Which Assistants Work Today

Assistant Supports a hosted MCP server Where you paste it
Claude (web, desktop, phone) Yes, and we tested ours here Settings, Connectors, Add custom connector
ChatGPT Yes Settings, Connectors, Create, with developer mode on
Claude Code Yes claude mcp add --transport http
Cursor Yes Settings, MCP, Add new server, Streamable HTTP
Gemini CLI, Zed, Windsurf Yes Their own MCP settings

Support moves quickly, and an assistant that does not take a custom connector today may take one next month. If yours is missing, the same tools are available as a local package, covered in the documentation.

What It Will Not Do

Being honest about the edges saves you an afternoon. The first one matters most, because it is the thing people get backwards:

  • The assistant is not what answers your customers. ChatGenius already does that around the clock, on our servers, whether or not you are online. An assistant only acts while you are talking to it, so it is how you inspect and steer the automation, never a replacement for it. Connecting Claude does not switch anything off.
  • It cannot reach personal Instagram accounts. Meta's API does not expose them, to anyone.
  • It will not act silently. Tools that send a message or publish something are flagged, so your assistant asks before using them. That is deliberate.
  • It needs the paid add-on. The MCP server is part of the ChatGenius Webhook and API add-on, $29 a month on the Creator plan and above. Full prices are on the pricing page.
One security habit worth having

Message text in your inbox was written by other people. An AI reading it is reading strangers' words, and a customer could in principle write something designed to look like an instruction. This is a known weakness of every AI assistant with tool access, not something specific to Instagram. It is also why the read-only option exists, why sending requires your approval, and why we tell the assistant that message text is data rather than orders. Start read-only for a week and see what you use.

Setting It Up

Three steps, about five minutes, assuming Instagram is already connected to ChatGenius.

  • 1. In your assistant, add a custom connector and paste https://sumgenius.ai/api/mcp. Leave whatever it detects on its own.
  • 2. It sends you to a ChatGenius page. Sign in as the account owner, choose read only or read and act, click Allow.
  • 3. Ask it something. "Who messaged me this week?" is a good first test.

The full instructions, including the local package for Claude Desktop and Cursor, are in the MCP documentation. There is also a REST API and webhooks if you would rather write code than talk to an assistant.

Is This Worth Doing?

If you already have DM automation running, yes, and the reason is not the automation. It is that asking questions of your own data in plain English is faster than any dashboard, including ours. Two questions alone justify the setup: "what are people asking that my AI cannot answer," and "why is that account growing faster than me?"

If you have no automation yet, connect an assistant read-only and spend a week asking it what actually happens in your DMs. You will learn more about your own business in that week than from any report. Then decide what to automate.

And if you were about to install a free Instagram MCP server that wants your password, do not. The account it risks is the one your customers reach you on.