openclip

OpenClip is an automated video processing tool based on Python, FFmpeg, and uv, primarily designed to identify and export highlight clips from long videos such as monologue-style recordings, live stream replays, interviews, and podcasts. After users input a video URL or local file, the workflow includes downloading or verifying the video, retrieving platform subtitles or performing local speech transcription, splitting long videos, AI analysis and summary of candidate segments, clip generation, subtitle file generation, thumbnail creation, and optional title addition and subtitle burning. Core capabilities include: - Support for Bilibili, YouTube video links, and local video files; Streamlit also supports browser file uploads. - Prioritizes platform subtitles for transcription; local ASR can route by language, using Whisper for English and Paraformer for Chinese, with fallback to Whisper when unavailable. - AI filters highlights based on dimensions like content, engagement, and entertainment value; users can specify focus topics via natural language using --user-intent, and can also add presenter background and custom prompt templates. - Configurable parameters including number of clips, duration presets, title style, font size, and thumbnail text position and color. - Automatically outputs independent video clips, SRT subtitles, summary Markdown, and thumbnail images. - Optional deep optimization mode adds AI review, boundary refinement, and secondary review after candidate segment aggregation to improve clip independence. - Optional subtitle hard-burning requires FFmpeg with libass; bilingual subtitles can be translated and burned using the selected LLM. - Preview speaker recognition can label speaker names based on reference audio, suitable for multi-person dialogue scenarios, requiring additional speakers dependency and HuggingFace Token configuration. - Post-processing Clip Editor supports adjusting individual clip boundaries, subtitles, and thumbnail titles, with optional speed-up re-rendering. The tool offers three interfaces: Streamlit web UI supports background tasks, multi-task concurrency, task persistence, progress tracking, Cookie mode configuration, and local network access; CLI is suitable for scripting; Agent Skills can be invoked by agents supporting skills, such as Claude Code, TRAE, and Cursor, enabling video processing tasks via natural language. AI interfaces support Qwen, OpenRouter, Zhipu GLM, MiniMax, and custom OpenAI-compatible APIs, compatible with services like LM Studio, vLLM, One API, and New API. Remote downloading supports three modes: no cookies, browser cookies, and Netscape-format cookies.txt; for YouTube, Deno or Node installation is recommended in some scenarios. Processing results are saved by default under the processed_videos directory, organized into subdirectories such as downloads, splits, clips, and clips_post_processed. The project is licensed under MIT.