Connect Claude to tldw.sh, send a YouTube link and describe what you need: the main argument, practical steps or an answer to a question.
tldw.sh retrieves available captions and saves the analysis. Claude submits the request, checks progress and retrieves the result when it’s ready.
Choose your Claude client
Claude web and Claude Desktop use the hosted MCP connector with OAuth. You sign in to your own tldw.sh account and approve access. You do not need to create an API key, run a terminal command or install the Claude Code plugin.
Claude Code is a separate coding assistant. It has a different plugin/MCP installation path, documented on Connect your agent. Installing a Claude Code plugin does not connect Claude web or Desktop.
Custom-connector controls depend on your Claude account and organization. See Claude's custom connector instructions if your interface differs.
Connect Claude to your account
-
Create a tldw.sh account and complete the one-time account verification. Eligible Free accounts receive 100 monthly credits for captions and text, with a 30-minute per-source limit. No payment card is required for Free.
-
In Claude web or Desktop, open Customize → Connectors → Add custom connector. Enter this server URL:
https://api.tldw.sh/mcp -
If Claude asks for authentication settings, choose Sign in now and Register automatically. This uses the dynamic client registration supported by tldw.sh. Do not select Claude's published-identity method: tldw.sh does not currently advertise support for that method.
-
Add the connector, choose Connect, and sign in to tldw.sh. Review the client identity and requested permissions before approving. No OAuth client secret or manually pasted API key is needed for this path.
-
Confirm the connection in Claude. Adding a URL and authorizing an account are separate steps.
The account verification happens in your browser. Later tool calls do not need a CAPTCHA.
Add your instructions
Enable the connector for your conversation and provide a public YouTube link with a specific request:
Use tldw.sh to analyze this YouTube video: [paste the link]. Explain the main argument, the evidence supporting it and the most useful practical takeaway. Keep it concise and avoid repeating points.
You can also ask for a particular output language or audience:
Summarize this tutorial for an experienced Python developer. Focus on implementation decisions and limitations. Write the result in Spanish.
Instructions change the generated analysis. They do not alter the saved original captions or create dubbed audio or translated subtitle files.
Claude calls analyze_media, checks progress with get_job or wait_job, then reads the result with get_summary. If a wait times out, ask Claude to resume the existing job instead of submitting the video again.
Retrieve the transcript separately
A summary is generated analysis. A transcript is the saved source text. Ask Claude to retrieve the saved transcript when you need to check which source wording supports an answer:
Get the original transcript for that completed analysis. Show the section supporting the second takeaway.
The get_transcript tool uses the completed job's summary ID. It reads the saved captions rather than starting another transcription or summary job. Caption wording, timing and language remain those of the source, including any caption errors.
Understand the coverage
Captions describe speech. They may miss charts, silent demonstrations or on-screen text. Automatic captions can contain errors.
Fresh audio transcription is unavailable on every plan. Missing captions fail without a processing-credit charge; stop or supply authorized source text. Private, removed, restricted or platform-blocked videos may fail. Processing time varies.
Browse saved transcripts
Each submitted link produces both a summary and an original transcript. Use get_summary and get_transcript with the same summary ID. Reading either saved result starts no new processing or model call.
Use list_transcripts for your latest 50 sources, next_cursor for more, UTC since/until filters or unread_only to find unread sources. Listing and reading source pages do not mark them read; call mark_transcript_read explicitly. The web Transcripts & text library opens source text and marks it read only after loading.
Manage access and usage
Caption processing uses 1 credit per source minute. Text is billed by estimated reading time. Connected agents share your account's credits. Reading a saved summary or transcript uses no processing credits.
You can revoke access in Settings → Connected apps. For the latest limits and error recovery, use the API documentation and current pricing.