Use Case
Instagram MCP Server: Your Tracked Instagram Data Inside Claude, ChatGPT and Codex
How an Instagram MCP server lets your AI read tracked accounts, saved posts, comments and metrics, and what still needs a quoted run.
By Bell Chen, founder of Superdirector — how the analysis works.
Reference visuals
Frames and thumbnails behind this guide
People search for an "Instagram MCP server" because they have already moved their thinking into an AI app and want Instagram to come with them. The Model Context Protocol (modelcontextprotocol.io) is the open standard that lets a client such as Claude, ChatGPT or Codex call tools on a remote server, so the question is less whether a connection is possible and more what the server on the other end actually holds.
That is where most expectations go wrong. An MCP server is not a window into all of Instagram. It is a set of tools over some store of data, and the answers your AI gives are only as good as that store. This page describes one concrete arrangement: an AI app connected to a Canvas that tracks the Instagram accounts you chose, with clear lines around what is saved, what costs a run, and what the AI cannot see.
What a connected AI can read, and what it cannot
A Canvas is a workspace with sources on one side and actions on the other. On Instagram, the sources are your own account, the competitor accounts you track, the inspiration accounts you keep for reference, and keyword and hashtag sources. Each completed read saves what it returned: posts with their captions and metrics, comments when comment collection was part of the read, and account-level metrics. Your notes, research and saved drafts sit beside them in Files.
Once connected, the AI reads all of that through native tools. It can page an inventory of what the Canvas tracks, read the exact saved posts, comments or metrics from one committed read, and search your saved notes. It returns no images: covers and other pictures are viewed on the Canvas in the web app. Each record carries the date it was observed, so the AI can tell you when a number was true.
Three things sit outside that. First, the AI does not browse Instagram live: fresh or wider data is a new collection run, priced in credits, and it starts only after you approve the quote. Second, saved posts are not watched videos; frames and transcripts come only from a video analysis report. Third, nothing is published. The AI drafts and saves, and you post.
Setting it up and asking the first useful question
| Stage | When / duration | Tools | Deliverable |
|---|---|---|---|
| 1. Add the server and approve access | about 5 minutes | your AI app settings, a browser | an authorized MCP connection |
| 2. Check what the Canvas already tracks | one prompt | account_status, list_canvases, knowledge_inventory | a list of tracked Instagram sources with their last read dates |
| 3. Ask from saved evidence first | a few minutes | read_source_evidence, context_search | an answer that cites saved posts, comments or metrics |
| 4. Quote and run when the saved data is not enough | minutes to queue, longer to complete | quote_module_run, run_module, get_module_run | a committed run with a run ID |
| 5. Save the output where the next chat can find it | one prompt | save_artifact, memory_append, open_canvas | a brief, script or plan in Canvas Files |
Add the server and approve access
Add the remote MCP endpoint in your AI app's connector or MCP settings and sign in when the browser opens. Authorization follows the protocol's OAuth flow (modelcontextprotocol.io), so the AI never sees your password, and the grant covers the Canvases your account owns, not anyone else's.
Codex and Claude Code take a remote HTTP server in their MCP configuration. ChatGPT and Claude on the web add it as a custom connector where the plan and workspace allow it. If the tools do not appear after sign-in, reload the conversation; a running chat does not always pick up new tools on its own.
Check what the Canvas already tracks
Ask the AI to list your Canvases and tell you which Instagram sources each one tracks. The inventory names every account by role, the keyword and hashtag sources, and how many posts each completed read retained. These checks start no provider work and cost nothing.
Read the dates. A competitor read from five weeks ago answers a different question than one from yesterday, and the AI should say which it is using.
Ask from saved evidence first
Ask the question you actually have: which recent posts from the three competitors drew the most comments, what the comments under them asked about. The AI reads the saved records page by page and reports what coverage it had, including posts whose comments were never collected.
Covers stay on the Canvas: the connection returns no images, so open the Canvas in the web app to compare framing, text overlays and color yourself. The AI can tell you which saved posts have a retained cover; a missing one is not fetched behind your back.
Quote and run when the saved data is not enough
If the question needs fresher posts or an account you do not track yet, the AI asks the Canvas for a quote. The quote names the credit price and expires after ten minutes. Nothing runs until you approve it in the chat; the AI then starts the run and polls it to completion.
An admission receipt is not a result. The AI should hold the run ID and report the outcome when the run commits, including anything it could not collect.
Save the output where the next chat can find it
Ask the AI to save the brief or plan to the Canvas. It lands in Files with version history, and you can open the Canvas beside the conversation to see it. The next conversation, in the same app or another one, starts from that saved work instead of from a blank page.
Common mistakes
Treating a connection as live Instagram access. The AI reads what was saved, with dates. If you ask about this morning and the last read was last week, a careful AI says so; a careless prompt can hide it. Ask for the observation date beside any number.
Reading a caption and calling it video analysis. A saved post tells you what was written and how it performed. It does not show the edit, the pacing or the spoken words. When those matter, run a video analysis report and read its result.
Approving runs without reading the quote. Collection across many accounts adds up. Ask the AI to show the aggregate scope and the current quotes before you approve anything, and approve one run at a time when you are unsure.
When you do not need an MCP server
If you look at one or two posts a week, the native Instagram app and its insights are enough, and copying a caption into a chat is faster than setting up a connection. The MCP route earns its setup when the work spans several accounts, repeats every week, or needs the comments and metrics in the same place as the drafting. It is also not a substitute for Instagram's own analytics on your account; it complements them with competitor and inspiration context.
Where Superdirector fits
Superdirector is the Canvas on the other end of this connection. Connecting, account status and Canvas inventory are free; paid plans start at $29 per month and new collection is metered in credits with a quote before every run. A verified account gets a one-time 20-credit MCP starter inside the AI app, enough for at most one account check and one X text draft. The MCP starter is not a trial of the plan; Starter has its own 7-day free trial. Your AI cannot post; posting is approved in the Canvas.
The same Canvas works across Instagram, TikTok, YouTube and X; for the multi-app picture see (Social Media MCP for ChatGPT and Codex: One Canvas, Four Platforms, Any MCP Client), and for a worked competitor review see (Competitor Analysis in Claude: Compare Tracked Accounts, Posts, Comments and Metrics).
Disclosure: I build Superdirector, the product described on this page. Tool names and limits are the ones the MCP server advertises today and can change; your AI reads the current capabilities before it promises an action. Instagram is a trademark of Meta, which is not affiliated with this page.