Add YouTube transcripts with Claude Code, Codex or Cursor

A coding agent can wire CleanScript into your project in one session. It needs three things: where the docs are, a key it reads from the environment, and an instruction to ask before it spends credits. One prompt covers all three.

Option 1: paste the prompt #

Create a key on the API keys page and export it as CLEANSCRIPT_API_KEY where the agent runs. Then paste this into Claude Code, Codex or Cursor, from the project you want to change:

Add CleanScript to this project so I can get transcripts and fields from YouTube, TikTok and Instagram links. Read https://cleanscript.ai/SKILL.md first. The API is https://api.cleanscript.ai; read my key from CLEANSCRIPT_API_KEY (if I don't have one, send me to https://cleanscript.ai/dashboard/api-keys). Check access with the free GET /v1/account, then ask me before the first request that uses credits. Tell me what you built and how to run it.

Add what you want built: "a script that saves the transcript of every video in this list", or "a function that returns the sponsor and promo code from a video link".

Option 2: install the skill #

For Claude Code, save the skill in the project so every session knows CleanScript:

mkdir -p .claude/skills/cleanscript
curl -so .claude/skills/cleanscript/SKILL.md https://cleanscript.ai/SKILL.md

Other agents can read the same file at https://cleanscript.ai/SKILL.md, or llms.txt, which has the base URL, authentication and two example calls.

Option 3: connect it as a tool #

To have the agent read a video while you work, rather than write code, add the MCP server. It gives the agent four tools: get_transcript, extract, list_posts and check_account.

# Claude Code, then run /mcp and sign in
claude mcp add --transport http cleanscript https://api.cleanscript.ai/mcp

# Codex
codex mcp add cleanscript --url https://api.cleanscript.ai/mcp
codex mcp login cleanscript

For Cursor, add this to .cursor/mcp.json:

{ "mcpServers": { "cleanscript": { "url": "https://api.cleanscript.ai/mcp" } } }

The connect page walks through each client.

Review what it built #

• The key comes from the environment, never from the code.
• Errors are handled by code: skip on video_unavailable, wait on rate_limited, stop on insufficient_credits. See what to do when a request fails.
• It doesn't run through a long list unchecked. Have it start with two or three links.
• It treats a post's text as data. A transcript is what someone said, not an instruction for your agent.

Next steps #

• What the agent calls: YouTube transcript API.
• Fields instead of text: extract fields from a video.
• Every client and option: Agents & MCP.