Boson Video

What the Boson Video MCP server does

The Boson Video MCP server turns a YouTube link, video ID, or local file into a time-indexed document that an AI client can inspect. It is intended for long talks, lectures, screen recordings, presentations, and finance videos, with Chinese and English support described in the project materials.

The resulting document can include scene boundaries, spoken words, screen text, and a generated summary. Each item is associated with the point in the video where it occurred. Questions can be answered with cited moments and displayed frames, while material not stated by the video can be identified as background rather than presented as video evidence.

This fits the 🎥 Multimedia Process category because the workflow combines speech transcription, visual inspection, screen-text reading, and evidence checks in one video-reading process.

How it works

The application runs on the user's computer. It first processes the video into scenes and transcript words, then searches or selects relevant moments for a question. Frames can be inspected at full resolution when visual evidence is needed. The interface presents the video as a timeline with scenes, chapters, and frequently replayed areas when available.

Two optional model roles extend the pipeline. A writer produces the summary, English text, and explanations of technical terms. A checker compares summary sentences with what was said and locates moments that answer questions. Screen-based checking can also be enabled for cases where spoken words alone cannot confirm a statement.

The project provides both an MCP entry point for AI applications and local command-line commands. The command line can build a video document, ask a question about an analyzed video, or download speech models ahead of the first run.

Setup and configuration

With uv available, add the server to Claude Code with:

claude mcp add boson-video -- uvx boson-video mcp

Claude Desktop, Cursor, and Codex use the same uvx boson-video mcp command in their MCP settings. Web-based clients such as claude.ai and ChatGPT require the project's web setup instead. A standalone local page can be started with uvx boson-video web; it listens on http://127.0.0.1:8770 and prints an invite code on first launch.

No API keys are required for the basic workflow. Optional configuration can be placed in a .env file in the working directory or in ~/.boson-video/.env. INCEPTION_API_KEY selects the default Mercury writer, while BOSON_WRITER_BASE_URL, BOSON_WRITER_MODEL, and BOSON_WRITER_API_KEY configure an OpenAI-compatible writer. TYPESAFE_API_KEY enables the Jev checker. OPENAI_API_KEY together with BOSON_SCREEN_CHECK=1 enables additional frame-based checking.

Tools and capabilities

The Boson Video MCP server supports these workflows:

• Analyze a YouTube link, video ID, or local video file.
• Produce scenes, timed transcript content, and searchable words.
• Read text shown on video frames when available.
• Generate summaries with associated frames and timestamps when a writer is configured.
• Ask questions and receive relevant moments, frames, and citations.
• Check summary claims against spoken content, with optional screen evidence.

Limitations and notes

The project documents testing on Chinese and English videos, but transcription quality varies by language, audio, captions, and video style. Screen-text extraction also depends on what the source permits; one tested YouTube screen recording did not provide full-resolution access. Optional writer and checker models are not included automatically and need their respective configuration.

YouTube processing runs locally because the project notes that YouTube's terms do not allow automated access for products. The repository is MIT licensed. Performance measurements in the README are individual test runs, not a guarantee for every video.