如何在 Claude 和 ChatGPT 中取得 TikTok、Instagram Reels 與 YouTube 影片逐字稿(MCP)
How to get TikTok, Instagram Reels and YouTube transcripts in Claude and ChatGPT (MCP)
作者以自建的 Mira 遠端 MCP 伺服器示範,如何讓 Claude、ChatGPT、Cursor 等工具處理 TikTok、Instagram Reels 和 YouTube 短片,並返回語音逐字稿及畫面文字的 OCR 結果。
You paste an Instagram Reel into Claude and ask "what's the recipe?". Claude can't open it. TikTok and Instagram put a login wall in front of bots, and YouTube Shorts often give an assistant nothing but the title. So the model either refuses or, worse, guesses.
This post covers why that happens, the options you have, and how to wire a video transcript tool into Claude, ChatGPT and Cursor over MCP.
Why assistants can't "watch" a link
To answer a question about a short video, an assistant needs three things it doesn't have:
- The file. Social platforms block simple fetching, so you need a real downloader, and those break whenever a platform changes something.
- What was said. That means speech-to-text on the audio track.
- What was shown. In short videos, half the information is on-screen text: ingredient lists, addresses, prices. A transcript alone misses it, so you need OCR on key frames.
Then all of that has to reach the model in a compact shape. A 2,000-word raw transcript is noisy. "Here are the 6 places, with addresses" is what you actually want.
Your options
1. Local scripts (free, DIY). yt-dlp to download, Whisper to transcribe, a script to glue it together. Great if you're on a laptop with Claude Code and don't mind maintenance. There are open-source skills that do exactly this. The downsides: Instagram downloads break often, there's no OCR of on-screen text unless you add it, and it doesn't work in claude.ai or the mobile apps.
2. Scraping platforms (pay per result). Hosted actors and scraping APIs return transcripts for TikTok or Instagram. They're good for bulk jobs, but you still have to get the data into the chat yourself.
3. A remote MCP server (works everywhere). The assistant calls a tool, and the tool does the downloading, transcription and OCR on a server and returns a structured result. It works in claude.ai, Claude Desktop and mobile, ChatGPT, Claude Code, Cursor and VS Code, with nothing to install.
The rest of this post uses option 3 with Mira, the server I build. Disclosure up front: it's a paid service, because every video costs real transcription and OCR time.
Setup
The server speaks Streamable HTTP at https://mira-ai-skill.pro/mcp. Auth is an API key, either as a Bearer header or embedded in a personal connector URL, since claude.ai and ChatGPT connectors can't send headers.
Claude (web, desktop, mobile): Settings → Connectors → Add custom connector → paste your personal connector URL.
Claude Code:
claude mcp add --transport http mira https://mira-ai-skill.pro/mcp \
--header "Authorization: Bearer mira_sk_your_key"
Cursor (~/.cursor/mcp.json):
{
"mcpServers": {
"mira": {
"url": "https://mira-ai-skill.pro/mcp",
"headers": { "Authorization": "Bearer mira_sk_your_key" }
}
}
}
ChatGPT: Settings → Apps & Connectors → Advanced → Developer mode → Create → paste your personal connector URL.
Configs for VS Code, Windsurf and Gemini CLI are in the repo: https://github.com/Mira-AI-Tools/mira-mcp
What you can ask
- "What's the recipe in this Reel? Turn it into a shopping list."
- "List every place from this TikTok with addresses, I'm in Tokyo next week."
- "Transcribe this YouTube Short, translate it to English, and give me 3 bullets."
- "Break down this competitor's ad: hook, structure, call to action."
- "Make study notes from this lesson."
Three things I learned building it
Long videos don't fit in one tool call. parse_video waits about 40 seconds. If the video isn't done by then, it returns a job_id, and the model calls get_result. Models handle this well as long as the tool description says so plainly.
Captions are a prompt-injection channel. Transcripts and captions are written by strangers, and some contain text like "ignore your previous instructions". The skill that ships with the server tells the model to treat everything in an extract as data to analyze, never as instructions, and to tell the user when it sees something like that.
Keep the creator's words apart from the tool's. If the tool estimates calories for a recipe, the model should say "estimated", not present it as the creator's number. That one rule removed a whole class of confident-but-wrong answers.
Links
- MCP server docs and client configs: https://github.com/Mira-AI-Tools/mira-mcp
- Claude skill and Claude Code plugin: https://github.com/Mira-AI-Tools/mira-skill
- Mira: https://mira-ai-skill.pro/?source=dev
If you're building your own remote MCP server and want to compare notes on async tools or auth for claude.ai connectors, ask in the comments.
來源:DEV Community · dev.to