Reference

MCP

Give Claude, Cursor, Windsurf or your own agent the ability to find and transcribe videos.

The MCP server gives an AI agent the same API as your key: it can find videos, transcribe them and read the results in the middle of a conversation. It charges your credits and honors your key's scopes. Nothing to install: point the client at the remote server.

Connect#

Claude Code

claude mcp add --transport http transcriptdock https://transcriptdock.com/mcp \
  --header "Authorization: Bearer td_live_YOUR_KEY"

Cursor, Windsurf, VS Code and other clients (mcp.json)

{
  "mcpServers": {
    "transcriptdock": {
      "type": "http",
      "url": "https://transcriptdock.com/mcp",
      "headers": { "Authorization": "Bearer td_live_YOUR_KEY" }
    }
  }
}

Local server#

For clients without HTTP transport (Claude Desktop, OpenClaw), run the same server locally with Node.js. The signed-in MCP page in your dashboard has this config with your key filled in.

mcp config

{
  "mcpServers": {
    "transcriptdock": {
      "command": "npx",
      "args": ["-y", "@transcriptdock/mcp"],
      "env": {
        "TRANSCRIPTDOCK_API_KEY": "td_live_YOUR_KEY",
        "TRANSCRIPTDOCK_BASE_URL": "https://transcriptdock.com"
      }
    }
  }
}

The remote server caps each tool call at about 40 seconds. A job still running after that comes back with its id; the agent can call wait_for_job again. Clients that only sign in with OAuth (Claude web, ChatGPT) are not supported yet.

Tools#

Tool	Does	Credits
`get_video_transcript`	URL in, transcript out. Submits, waits, returns text (or srt, vtt, segments).	1, or 2 per minute with AI transcription
`transcribe`	Submit a job and return it without waiting.	Same
`wait_for_job`	Wait for a job to finish.	Free
`get_job`	Job status, error and billing.	Free
`get_transcript`	Read a transcript by id as segments, text, srt or vtt.	Free
`list_jobs`	Recent jobs.	Free
`search_youtube`	Search YouTube.	1 per page
`list_channel_videos`	A channel's videos, newest first, with paging.	1 per page
`get_channel_latest_videos`	A channel's latest 10 videos.	1
`search_channel_videos`	Search inside one channel.	1 per page
`list_playlist_videos`	A playlist's videos.	1 per page
`list_tiktok_user_videos`	A TikTok profile's latest videos (up to 10) and profile stats.	1
`get_video_discovery`	Read a discovery by id, e.g. after a timeout.	Free

get_video_transcript#

The tool an agent should reach for when asked what a video says.

urlstringRequired: YouTube or TikTok video URL.
mode"captions_only" | "auto" | "transcribe"Optional: Default captions_only (1 credit). auto falls back to AI transcription when the video has no captions; transcribe always uses it (2 credits per started minute, paid plans).
format"text" | "srt" | "vtt" | "segments"Optional: text (default) is the most compact for an agent; segments returns timed JSON.
timeout_secondsintegerOptional: Default 120. On timeout the job state and id are returned instead.

call

{ "url": "https://www.youtube.com/watch?v=dQw4w9WgXcQ", "mode": "auto" }

returns

Never gonna give you up
Never gonna let you down
…

The same video with the same mode is answered from your workspace's stored result, free.

transcribe, wait_for_job, get_transcript#

The three-step version for agents that want control: transcribe takes url, mode, optional language and max_credits, plus an idempotency_key (any stable 8 to 128 character string, so a retried call never submits twice) and returns the job. wait_for_job takes job_id and an optional timeout_seconds. get_transcript takes the job's result_id as transcript_id and a format.

Finding videos#

The search and list tools return videos with url, title, duration and view count, ready to pass to get_video_transcript. channel accepts an @handle, channel id or channel URL; playlist an id or URL; user a TikTok @user or profile URL. YouTube tools page with limit (default 20) and cursor. TikTok returns at most 10 videos and no further pages; include_details adds likes, comments, hashtags and caption languages per video.

Errors#

A failed call returns isError: true with the API error code, message and request id, so the agent can explain it or change course. The codes are the ones in the API reference.

Troubleshooting#

Unauthorized: the key goes in the server config (header or env), not in your shell. Check it has not been revoked.
Tool not found: restart the client after editing its config.
Local server will not start: install a current Node.js LTS; npx comes with it.
NO_CAPTIONS: the video has no captions. Ask for mode: "auto".

Site menu

Documentation

MCP