Turning Every AI Video Summary Into Something You Can Actually Verify

ClipScript is an MCP tool that gives Claude, ChatGPT, and Gemini direct access to a YouTube video's real transcript — not just their own paraphrased summary of it.

• Whether it's researching an interview or following a tutorial, sometimes you need the exact transcript — the precise quote, the exact step — not a paraphrase. Get it wrong and you misquote a source or make a mistake building something.
• Every AI assistant offered the same shortcut: a summary in its own words, no real transcript behind it, no way to verify a claim or pull an exact quote — because LLM chats won't pull a full transcript directly; copyright restrictions block it.
• The only workaround was manual: open each video by hand, copy and paste the captions, clean it up before it was usable — repeated every time, across every video.

Real Transcript Data, Not a Paraphrase — Built as an MCP server (Python SDK + youtube-transcript-api) that pulls YouTube's actual caption data, so the model works from the real transcript instead of inferring one. yt-dlp handles metadata and chapters separately, since the transcript library doesn't expose those — a deliberate two-library split.

From Wall of Text to Something Usable — The first working version just dumped the entire transcript verbatim — unusable for a 45–60 minute video. Next came the summarize/outline/follow-up-question layer, then chapter-aware sectioning using YouTube's own chapter markers (falling back to ~30-second buckets when a video has none) — all while keeping the full verbatim transcript live underneath, so an exact quote is always one question away.

Fixing a False-Positive Security Flag — Claude Desktop kept flagging the tool's own formatting guidance as a suspicious embedded instruction. Two fixes failed — rewording the phrasing, then moving it into the tool's trusted docstring metadata. What worked was reframing the guidance as a non-mandatory suggestion instead of a command.

Making Timestamps Actually Clickable — Turning timestamps into markdown links that jump to the exact second sounds simple; the first attempt failed because the model paraphrased sections and dropped the link syntax in the process. Fixed by rewriting the guidance to explain why a link only stays clickable when copied exactly — not just what to output.

• No more manually copying and cleaning captions before I can work with them — every summary now comes with the real transcript behind it
• Exact quotes and clickable timestamp links for any claim, so I can verify or jump straight to the moment instead of re-watching
• Built for the moments that matter most: research, tutorials, and case studies — anywhere the actual wording, not someone's paraphrase of it, is what's needed
• Works across Claude, ChatGPT, and Gemini
• Available on GitHub right now — heading to the MCP marketplace next, so anyone will be able to install it

ClipScript is available on GitHub right now — I'll update this case study once it's live on the MCP marketplace. Either way, every video summary it gives you comes with the receipts: the real transcript, the exact quote, the exact moment.