bibi

bibi (BibiGPT) is an AI video and audio summarizer and repackager. It summarizes content from YouTube, Bilibili, podcasts, TikTok, Twitter/X, Xiaohongshu, and other online video or audio sources, and can optionally turn the takeaway into a TikTok-style vertical music video. It solves the problem of quickly understanding long-form media by producing summaries, chapter-by-chapter breakdowns, and transcript/subtitle extraction, and by remixing long videos into short vertical MVs. It works through the bibi CLI on macOS or Windows, or via the OpenAPI HTTP interface on Linux or any platform without the CLI.

The skill begins with an environment check via scripts/bibi-check.sh, which detects whether CLI mode or OpenAPI mode is available. CLI mode authenticates through desktop login or a BIBI_API_TOKEN, while OpenAPI mode uses BIBI_API_TOKEN only. It includes a China payment fallback: when neither auth is set, BibiGPT may return HTTP 402 Payment Required with a Payment-Needed header (the AI 收 / Alipay protocol), detectable via the stable stderr marker line [HTTP/402 Payment Required], which routes to references/billing-aipay.md. An intent-routing table maps requests to workflows such as quick-summary, deep-dive, transcript-extract, article-rewrite, video-to-tiktok-mv, batch-process, research-compile, export-notes, visual-analysis, account-check, library-browse, channels-manage, feed-latest, collections-manage, notes-manage, and advanced-tools. Direct CLI operations include bibi summarize with flags like --chapter, --subtitle, and --json, plus bibi auth check, bibi me, and bibi commands, with progressive help discovery.

It is designed for users who want fast video/audio summaries, transcripts, chapter summaries, note exports (Notion, Obsidian), or short-form video remixes, and it supports multilingual triggers including Chinese phrases. The CLI accepts local file paths directly (e.g. bibi summarize "/path/to/video.mp4"), while API mode requires uploading files to a publicly accessible URL first. When a request matches multiple workflows it asks one clarifying question, and when it matches none it asks the user rather than guessing.