Instagram reels research with an AI agent (MCP and skills)

November 8, 2026

An AI agent with live data access can do in ten minutes what used to take an analyst an afternoon: find reels on a topic, pull their numbers and transcripts, and summarize what works. The trick is giving it the right tools and asking in a way that keeps it honest.

How it works

The Model Context Protocol (MCP) lets an agent such as Claude Code, Cursor or Codex call external tools. A social data MCP server exposes tools like "search reels" or "get transcript". The agent decides which to call, reads the JSON, and writes the answer. Skills are instruction files that teach the agent a workflow, such as how to page through results or how to vet a creator.

Setup

1. Get an API key. ScrapeIO gives 100 free credits on signup.
2. Add the hosted MCP server, https://scrapeio.com/mcp, to your agent. The exact steps for each client are on the agents page.
3. Optionally install the skills: npx skills add xhoantran/scrapeio-agent.

Instagram tools the agent will see include instagram_reels_search, instagram_user_reels, instagram_post, instagram_comments, instagram_transcript, instagram_profile and instagram_similar_accounts.

Five prompts that work

1. Topic scan.

Search Instagram reels for "protein coffee" from the last month.
Return the top 15 by plays as a table: handle, plays, likes, comments,
date, URL. Then say what the top 5 have in common.

2. Account deep dive.

Get @brand's profile and last 30 reels. Give median plays, plays per
follower, and the 3 reels most above median, with links.

3. Hooks.

For the top 10 reels above, get transcripts and quote the first
sentence of each. Group them by hook type.

4. Audience questions.

Pull comments on the top 3 reels. List the 10 most repeated questions
with a count and one example quote each.

5. Lookalikes.

Find accounts similar to @creator, get each profile, and keep those
between 10K and 200K followers. Table with links.

Keeping the agent honest

Agents are good at summaries and occasionally bad at numbers. These rules help:

• Ask for links. Every claim about a reel should come with its URL so you can check it.
• Ask for the data table first, then the interpretation. You can see if the summary matches the rows.
• Ask for medians, not averages, and the sample size.
• Set a budget. "Use at most 100 credits" keeps a curious agent from paging forever. Each response reports credits_used, so the agent can track it.
• Spot-check. Open three linked reels and compare plays with the app.
• Date it. Ask the agent to state when the data was pulled.

What an agent should not do

Read-only data tools cannot post, like, follow or message, and that is a feature. Keep agents on research. Do not ask an agent to compile personal details about individuals; ask for patterns across accounts instead.

Cost of a typical session

At the larger pack price that is about 4 cents. Repeat calls within the cache window are free, so re-asking a follow-up question costs little.

Without MCP

If your agent cannot use MCP, it can still call the REST API with curl or a short script. Give it the base URL, your key as an environment variable, and the docs. The answers have the same shape either way.

Doing it with ScrapeIO

The agent calls the instagram reels search API, instagram transcript API and instagram comments API for you. Our Claude Code, Cursor and Codex guide has the client setup, and the same tools work on TikTok, YouTube and 30 more platforms.