Video MCP server
Let Claude, Cursor and Codex read a video: transcripts, summaries and timestamped moments. No key needed to start.
https://scribiz.com/mcpInstall the video MCP server
Pick your client and copy the snippet. Your own model reads the results and answers. Full install guide
- Claude Code
- Claude Desktop
- Cursor
- Codex
- VS Code
Nothing to install. Works from any machine.
No key. Captions, videos already processed, and one free video read (up to 2 hours). No on-screen notes.
$ claude mcp add --transport http scribiz https://scribiz.com/mcpWith an API key. Listen and Watch on new videos, using your minutes.
$ claude mcp add --transport http scribiz https://scribiz.com/mcp --header "Authorization: Bearer $SCRIBIZ_API_KEY"Four MCP tools for video
All four are read only. New reads can use the free allowance, a caption credit or minutes. Reusing a stored result is free; generating a new summary uses minutes.
get_video_contextmay read a new videoThe outline of a video: summary, chapters and key moments with links.
get_transcriptmay read a new videoThe transcript or one stretch of it, paged, as text, Markdown, SRT, VTT or JSON.
search_videomay read a new videoWhere a word or topic comes up, as ranked moments with timestamps.
get_jobfreeThe state of a long job, by its id.
YouTube caption tools, with a key
BulkExisting captions only, with the source timing kept. No model and no minutes: they use the separate caption credits. Caption workflow
preview_youtube_captionsList the videos of a playlist or channel and the most a batch can cost, before spending anything.
start_youtube_caption_batchStart a batch of up to 1,000 YouTube videos: one caption credit per new video, nothing for failures.
get_youtube_captionsRead one video of the batch, with the source timing kept.
get_youtube_caption_batchThe state of a batch, with honest per-video failures.
Details: parameters and output shapes, limits and cost and why transcripts are untrusted text.
No key or an API key
Remote, no key
Paste the URL and go. Captions when our server can read them, videos Scribiz already holds, and otherwise a model reads the link (times approximate, up to a minute off on a long video, no speaker labels). 1 free video, up to 2 hours. No on-screen notes.
Remote, with an API key
Listen, Watch and Both on any public video, from your minutes. The key goes in an Authorization header as a Bearer token and is the same key the API takes. Keys are made in the dashboard, after you sign in.
Stored on-screen notes come with every result at no cost, and watch: true on get_video_context reads the picture once, from your minutes.
Local, scribiz mcp
One command, after npm install -g scribiz and scribiz login (or scribiz setup): claude mcp add scribiz -- scribiz mcp. It runs on your machine, with your free Scribiz account (30 minutes a month) or your own Gemini key, which uses no Scribiz minutes, and it fetches links from your own connection. It takes the watch argument and serves stored on-screen notes, like the remote server with a key. Local files work from folders you allow. Links to your own machine and to private networks are refused unless you pass --allow-private-network. Tested on macOS only.
OAuth sign-in: For claude.ai connectors. The URL above also works without signing in.
Watch the video MCP server in 48 seconds
Real calls to the server, with the answers it gave. No sound. An agent reads a long video in three steps:
- 1
get_video_contextSummary, chapters and key moments with links, in under about 2k tokens. - 2
search_videoThe moments where a word or topic comes up, ranked, with timestamps. - 3
get_transcriptThe exact passage around a hit, by from and to.
Questions & answers
What MCP is, which clients work, what is free and what leaves your machine.
- Model Context Protocol, an open standard that lets an AI app call tools. Scribiz runs an MCP server, so your agent can ask for a video instead of you pasting a transcript into the chat.
- Claude Code, Claude Desktop, Cursor, VS Code and Codex, through a connection URL or the local command. claude.ai takes the URL as a custom connector. We have not tested ChatGPT yet, so we do not list it.
- You do not need a key to start. Without one, the remote server uses captions when it can read them, answers at once for videos Scribiz has already processed, and otherwise has a model read the link (1 free video, up to 2 hours). It never looks at the picture, so there are no on-screen notes. Looking at the picture needs an API key and uses minutes.
- Most video MCP servers read YouTube captions and stop when there are none. When there are none, Scribiz has a model read the video from its link. It handles links beyond YouTube, and answers from a compact summary before it ever sends the transcript. Reading the picture needs an API key.
- It can try. A transcript is text from a third party, and a video can say anything. Scribiz marks the content as untrusted data, wraps it in a labelled block and says so in the tool description. Treat it like any web page your agent reads, and keep tools that write or send behind a confirmation.
- The link or the audio goes to Scribiz and then to Google for processing. Media is deleted when a job ends.
Give your agent the video
https://scribiz.com/mcp

