MCP server

Assistants cannot watch video. Ask one about an Instagram link and it will either refuse or answer from the caption, which is usually wrong.

The MCP server fixes that. It gives the assistant one tool: hand it a Reel link, get the spoken words back. From there the model can summarise it, pull the claims out of it, or write from it.

What MCP is

Model Context Protocol is a standard way to hand tools to an assistant. Claude Desktop, Claude Code and a growing number of other clients speak it. You add a server once and every conversation can use it.

Install

You need an API key. Email hi@instagramtotranscript.com and ask.

Claude Code

claude mcp add instagram-transcript \
  --env ITT_API_KEY=sk_live_your_key_here \
  -- npx -y @instagramtotranscript/mcp

Claude Desktop and other clients

Add this to your MCP config file.

{
  "mcpServers": {
    "instagram-transcript": {
      "command": "npx",
      "args": ["-y", "@instagramtotranscript/mcp"],
      "env": {
        "ITT_API_KEY": "sk_live_your_key_here"
      }
    }
  }
}

Restart the client. The tools appear in the tool list.

The tools

transcribe_reel

Takes a public Instagram link. Returns the transcript, the detected language, the duration and the account handle.

Argument	Type	Required
url	string	yes
language	string	no
withTimestamps	boolean	no

Set withTimestamps when the model needs to point at a moment in the video. Leave it off for summarising, where the timings only add noise.

transcribe_many

Takes up to twenty links and returns one result per link. Failures come back marked rather than throwing, so one private post does not lose the whole batch.

Use this for competitor sweeps. It is far cheaper in tokens than twenty separate calls.

Things worth asking it

• "Transcribe this Reel and pull out every claim as a bullet list."
• "Read these five Reels and tell me what they all agree on."
• "Transcribe this and draft a newsletter opening in my voice."
• "What question does this Reel answer? Write three headlines for it."

Limits

The same rules as the API. Public posts only, a duration cap, and a rate limit set per key.

Private posts come back as an error the model can read, so it will tell you the post was unreadable rather than inventing a transcript.

Without a key

The free tool does the same job one link at a time in a browser. Paste a transcript into your assistant by hand to see whether the workflow is worth automating before you ask for access.