Turn YouTube videos into text transcripts in the chat

The agent talks not to YouTube itself but to api.supadata.ai, and takes four things from it: the transcript of a video, video and channel metadata, the video lists of a channel and a playlist, and search with filters by recency, duration and sorting. That is what the analysis is built from: what is being said in a video, how it compares with the rest on the channel and what comes out on the topic over the period you need.

The main fork here is where the text came from. Some videos have native subtitles, and then the transcript is returned as is; if there are none, AI transcription kicks in, which costs more and is billed per minute rather than per request. The agent checks the native track first and only then falls back to AI, and it states in the answer which of the two paths worked: recognised text has a different accuracy from author-made subtitles.

There is nowhere to write through Supadata: it only reads public YouTube data — uploading a video, leaving a comment or changing a channel through it is not possible. So the agent works here as a researcher: it collects transcripts and metadata, translates a transcript into the language you need, brings the videos of a channel or playlist into a single table and summarises what it found on the topic. Nothing in anyone's channel changes along the way.

Usage is counted in plan credits: 100 requests a month on the free plan, 1 credit for a transcript or metadata, 2 credits per minute for AI subtitles, and batch processing of a channel only on paid plans. A long video is therefore priced by its running time up front and the agent says what the transcription will cost instead of spending the remainder silently. If a video is unavailable or the credits have run out, it says so directly instead of returning empty text.