watch-and-answer

watch-and-answer lets an AI agent actually watch a video and answer questions about it. Paste a YouTube link, ask a question, and the agent answers from what the video shows and says, with timestamps.

It downloads the video with subtitles, then picks a route that fits each model's limits: send the raw video, transcode it, send a compact low-frame-rate copy by URL, or fall back to sampled frames plus transcript for long videos. It works as a native Hermes Agent tool and as a CLI for Codex, Claude Code and other agents.

Tech: Python, yt-dlp, FFmpeg, multimodal LLMs