My coding agent kept choking whenever a task needed what was inside a video or audio link. "Summarize this YouTube talk", "pull the transcript from this podcast episode", "what resolution is this clip" - it would either make something up or try to shell out to a downloader that immediately got blocked.
Two reasons this is hard for an agent:
- Agents run in cloud sandboxes, and those run on datacenter IPs. YouTube and TikTok block datacenter IPs, so even if the agent installs a downloader, the fetch fails with a bot check.
- You don't really want your agent installing and babysitting a media downloader plus ffmpeg just to read one link.
I build importly.io, and it has a hosted MCP server for exactly this. You point your agent at it once and it gets tools to download, transcribe, and read metadata from any video or audio URL. The fetching happens on Importly's side, so it works from any sandbox, and there's nothing to install. Here's the whole setup.
Point your agent at the server
The server URL is https://mcp.importly.io/mcp.
Claude Code:
claude mcp add --transport http importly https://mcp.importly.io/mcp --header "Authorization: Bearer your_key" Grab your_key from the API key page. New accounts get $3 in free credits, no card.
Cursor - add to .cursor/mcp.json:
{ "mcpServers": { "importly": { "url": "https://mcp.importly.io/mcp", "headers": { "Authorization": "Bearer your_key" } } } } Claude.ai or ChatGPT - add https://mcp.importly.io/mcp as a custom connector / custom MCP server and sign in with OAuth. No key to paste.
Then just ask
Transcribe https://www.youtube.com/watch?v=jNQXAC9IVRw and summarize it in one sentence.
The agent calls transcribe_media, gets back a transcript with timestamped segments, and summarizes it. If the video has no captions, it can retry with speech-to-text.
The tools it has:
-
import_media- download the media to a hosted link -
transcribe_media- transcript with timestamped segments -
get_metadata/get_basic_metadata- formats, duration, uploader, thumbnails -
check_balance- credits left
Every job is async: the submit tool returns a jobId and the agent polls the matching check_ tool. The tool responses tell it how, so you're not wiring any of that yourself.
What it costs
A caption transcript is $0.01. A 10-minute podcast with no captions is about $0.30 with speech-to-text. You can cap spend per job with maxCostDollars, so a surprise 3-hour stream can't run away with your credits. Like I said, new accounts get $3 free to try it.
Going from "works in the agent" to "works in the app"
Once the feature works in the chat, the same API key works in your app over plain REST - submit a job, poll for the result. Transcribe looks like this:
curl -X POST https://api.importly.io/transcribe \ -H "Authorization: Bearer your_key" \ -H "Content-Type: application/json" \ -d '{"url": "https://www.youtube.com/watch?v=jNQXAC9IVRw"}' # -> { "jobId": "95be812f-...", "status": "queued" } curl "https://api.importly.io/transcribe/status?id=95be812f-..." \ -H "Authorization: Bearer your_key" Or just hand it to the agent that already has Importly connected: "add import-from-URL to our upload form". It pulls the working code for your stack from the get_integration_guide tool and writes it.
That's it. If you've been fighting bot checks on a server or watching your agent fail on video links, this is the shortcut. Full docs are at importly.io/docs/integrations/mcp.


