Watch Skill: local-first video indexing pipeline for LLMs via MCP

The author has released Watch Skill, an open-source project that creates a local-first video indexing pipeline for Large Language Models. The system analyzes videos once to extract transcripts, OCR data, scene boundaries, and representative frames, allowing LLMs to retrieve evidence without reprocessing the video.

• Hybrid retrieval combining full-text search (FTS) and embeddings.
• Scene detection used instead of fixed frame sampling.
• Timestamp-backed retrieval ensuring answers link back to the original video.
• All data stored locally so indexed videos do not leave the machine.
• Exposure via MCP, CLI, and REST API for model swapping.

The approach positions the LLM as a reasoning layer rather than a video-processing layer, enabling users to swap between different local models without changing the indexing pipeline.