Scribiz

What Scribiz MCP server does

Scribiz MCP server connects an MCP-compatible AI client to public video links. The hosted service accepts videos from YouTube, TikTok, Instagram, X, and Vimeo, then returns context an agent can use without downloading a file from the local machine. Results can include the title, duration, summary, chapters, key moments, transcript text, and links that open the source video at a relevant timestamp.

The service is designed for targeted reading rather than sending an entire transcript to the model. An agent can first request an overview, search for a phrase or topic, read only the relevant transcript section, or ask a direct question and receive three to five cited moments.

How it works

The hosted endpoint is available at https://scribiz.com/mcp and uses HTTP transport. A typical workflow starts with get_video_context, followed by search_video, get_transcript, or ask_video depending on the task. If processing is still underway, get_job can check the video job and retrieve its status.

All five tools are read-only and do not modify files. Results identify which processing layers were used, such as captions, listening, watching, or model-generated summaries, along with the minutes consumed. Without an API key, the service can use readable captions and previously processed videos. For videos without usable captions, it may read the public video; those timings are approximate, speaker labels are unavailable, and image-based details are not included.

Scribiz MCP server wraps video-derived text as untrusted content and marks results with untrusted_content: true. Agents should treat transcript text as data rather than instructions, especially in sessions that can run commands, write files, or send messages.

Setup and configuration

Add the hosted server URL to an MCP client. Claude Code can use claude mcp add --transport http scribiz https://scribiz.com/mcp. Claude Desktop and claude.ai support a custom connector with no sign-in, while Cursor, VS Code, and Codex use their respective MCP configuration files. The repository also provides a Claude Code plugin that installs the server and a usage skill; use the plugin or the direct server setup, not both.

An optional Scribiz API key is passed as a bearer token. Authenticated requests enable watch: true for get_video_context or ask_video, allowing stored or newly generated notes about on-screen content. The hosted service accepts public URLs only and cannot read local files.

A separate local command-line installation is available through npm. It requires Node 24, ffmpeg, and yt-dlp, and the documented testing is on macOS. After login, it can be registered as a local MCP process.

Tools and capabilities

• get_video_context: Returns title, length, summary, chapters, key moments, and timestamp links.
• search_video: Finds ranked moments where a word or phrase is spoken or shown.
• ask_video: Answers a video question with three to five cited moments.
• get_transcript: Returns transcript pages in TXT, Markdown, SRT, VTT, or JSON.
• get_job: Checks a video that is still being processed.

Limitations and notes

The no-key connection has daily allowance limits, including about 10 minutes of video reading per connection, videos up to 15 minutes, five questions, and 30 caption lookups. Shared limits may also apply. Previously processed videos and accessible captions continue to work in some limit cases, and responses report usage and reset timing.

The hosted service sends the video link and questions to Scribiz. Depending on the processing path, captions, audio, or video data may be sent to Google for transcription or analysis. Nothing from the local machine is uploaded by the hosted setup. Review the provider's privacy policy before using sensitive material.