YouTube Transcript MCP（jkawamoto）

Quick Verdict

YouTube Transcript MCP is an unofficial community server maintained by jkawamoto. It retrieves transcript tracks and video metadata that are already available through YouTube. The current version reviewed here is 0.7.0, released under the MIT license for Python 3.10 or newer. It exposes four tools: get_transcript, get_timed_transcript, get_video_info, and get_available_languages. That scope makes it useful for research, summarization, translation, and timestamp lookup inside an MCP client.

The most important limitation is easy to miss: this is not an automatic speech recognition system. It does not listen to a video’s audio and create a new transcript. If YouTube does not expose an accessible manual or automatic caption track, the core transcript tools cannot manufacture one. Installation sources also differ materially. The current GitHub source and the .mcpb asset attached to release v0.7.0 represent the reviewed feature set, while PyPI remains at the older 0.3.5 package. Users should not assume that a plain PyPI install includes the current four tools or URL support.

The software itself is free, but access, rights, inference cost, and data handling remain the operator’s responsibility. Transcript text can contain instructions aimed at an AI model. Treat all retrieved text as untrusted content: it must not override system rules, authorize actions, or cause an agent to invoke privileged tools. Because the server depends on YouTube access, this page marks needsVPN as true while making no recommendation about network services.

Best For

This server is best for people who already use an MCP-compatible client and need searchable text from public videos that have captions. A researcher can list available languages, retrieve timestamped transcript pages, and ask a model to organize claims while preserving a path back to the original video. A student can locate a term in a long lecture and jump to the relevant segment. Editors can use it for preliminary review before watching and verifying the source.

For multi-source research where videos are only one input among documents and notes, NotebookLM may provide a more complete research workspace. For real-time meeting transcription, speaker-oriented notes, or audio that has no existing captions, Otter.ai addresses a different requirement. This MCP server is not a good fit for private videos requiring an authenticated account, videos without accessible captions, speaker diarization, high-accuracy legal transcription, or large-scale harvesting. It should also be avoided when a team cannot establish transcript rights, review model-provider data terms, or isolate untrusted text from high-authority agent tools.

Key Features

• Four read-oriented tools: plain transcript retrieval, timestamped transcript retrieval, video metadata, and available-language discovery have separate interfaces.
• Language selection: get_transcript and get_timed_transcript accept a lang value and request English by default when no language is supplied.
• Long-response pagination: transcripts above the response threshold are split and returned with a next_cursor; operators can also lower the response limit.
• Several YouTube URL forms: version 0.7.0 added /shorts/, /embed/, and /live/ handling alongside standard youtube.com/watch?v= and youtu.be/ links.
• Metadata retrieval: get_video_info gives the client contextual video information, although it is not a replacement for every field in the official YouTube Data API.
• Language discovery: get_available_languages lets a workflow inspect actual caption tracks before requesting a language that may not exist.
• Current installation choices: operators can run uvx --from git+https://github.com/jkawamoto/mcp-youtube-transcript mcp-youtube-transcript, install the v0.7.0 .mcpb, or use the documented Docker image.
• No YouTube API key requirement: the project uses youtube-transcript-api and yt-dlp, but requests still depend on YouTube behavior, regional availability, and IP controls.

The four tools are read-oriented, which is narrower than a server that uploads, edits, or deletes account content. Narrow does not mean risk-free. The process performs outbound requests, returned content enters the MCP client, and a later model request may send that content to a separate provider. Each stage needs its own policy.

Use Cases

A careful workflow starts with a supported URL such as https://www.youtube.com/watch?v=..., https://youtu.be/..., or the current Shorts, embed, and live path forms. Call get_available_languages first, then request either plain or timestamped text in a language that exists. Follow next_cursor for long videos and store the video URL, selected language, retrieval date, and timestamps with any notes. This preserves provenance instead of retaining only a model-generated summary.

Course and interview search is another practical use case. A model can identify where a subject appears in the transcript, but the user should open the video at those timestamps and verify the spoken and visual context. Automatic captions often misrecognize names, numbers, accents, and technical terms. Charts, demonstrations, code shown on screen, and non-verbal events may not appear in the transcript at all. For workflows that need transcription from source media plus manual correction and editing, Descript is a more relevant comparison.

The server should not be treated as a bulk acquisition layer. Rate limits, IP blocks, region restrictions, removed videos, and upstream HTML or API changes can interrupt requests. A production workflow needs timeouts, bounded retries, concurrency limits, and a failure state that does not invent missing text. If a transcript is unavailable, the model should report that fact rather than infer content from a title or metadata.

Prompt injection deserves an explicit control. A video or caption author can include text such as “ignore previous instructions” or ask the reader to reveal secrets and call another tool. The transcript server merely returns that text, but an agent may interpret it as an instruction. Keep retrieved captions in a quoted data field, state that they are content to analyze rather than commands, disable unrelated filesystem or messaging tools, and require human approval for any external side effect.

Pricing

The repository and release v0.7.0 are MIT licensed and do not charge a software subscription. Operation can still consume network resources, local compute, and paid model tokens in the MCP client. Long transcripts are expensive context. Pagination prevents a single oversized response, but every page summarized or analyzed may generate another inference call. Set maximum video duration, cursor count, concurrency, model budget, and retention limits before automating a collection of URLs.

Version provenance matters more than the zero-dollar license. GitHub v0.7.0 was released in July 2026 with four tools, expanded URL handling, and an .mcpb asset. PyPI still reports 0.3.5 and therefore does not represent the current repository. A Git-based uvx configuration should pin the v0.7.0 tag or an exact commit instead of following main indefinitely. Claude Desktop users choosing the bundle should obtain it from the project’s GitHub release, confirm the displayed version, and avoid third-party repackaging.

There may also be content and compliance costs. Public accessibility does not automatically grant permission to reproduce, redistribute, translate, train on, or commercially package a transcript. Retain attribution and source links, quote only what the use case permits, and review YouTube terms plus the creator’s rights. A model provider may have its own retention and training controls, so sending an entire transcript for analysis is a separate data decision from fetching it locally.

Pros

• Four clearly separated tools cover transcripts, timestamps, video information, and language discovery.
• Version 0.7.0 accepts common watch, short-link, Shorts, embed, and live URL forms.
• No YouTube Data API key is required for basic use.
• Cursor-based pagination gives operators a way to bound long transcript responses.
• MIT licensing and Python 3.10+ make the source relatively straightforward to inspect and run.
• A current .mcpb release asset provides a convenient Claude Desktop installation path.
• Read-oriented scope is smaller than that of servers capable of changing channel or account data.

Cons

• It reads existing caption tracks and provides no ASR fallback when captions are absent.
• It is a community project, not an official YouTube integration, and upstream changes can break retrieval.
• PyPI 0.3.5 is stale relative to GitHub 0.7.0 and can create confusing tool and URL differences.
• Automatic captions can contain significant errors, while important visual information is omitted entirely.
• YouTube reachability, regional controls, account restrictions, and IP blocking affect reliability.
• Transcript data enters the MCP client and may then be sent to the selected model provider.
• Captions can carry prompt-injection text that becomes dangerous when an agent has privileged tools.
• Saving, transforming, or redistributing captions still requires platform and rights analysis.

Alternatives

FAQ

Does this server perform speech-to-text or ASR?

No. It requests caption tracks that YouTube already exposes, whether those tracks were created manually or automatically. If a video has no accessible captions, the server cannot listen to the audio and create a replacement transcript.

Should I install the PyPI 0.3.5 package for the current feature set?

No. This review covers GitHub v0.7.0, while PyPI remains at 0.3.5. Use a pinned v0.7.0 Git tag or commit, or the .mcpb attached to the official v0.7.0 GitHub release, if you need the reviewed tools and URL handling.

Which YouTube URL formats are supported?

The current version supports normal youtube.com/watch?v= links and youtu.be/ short links, and v0.7.0 added /shorts/, /embed/, and /live/ forms. Private, deleted, login-restricted, region-blocked, or captionless videos can still fail.

Can I freely save and republish any public transcript?

Do not make that assumption. Visibility is not a waiver of copyright or platform terms. Preserve provenance, limit quotations, review YouTube’s terms and the creator’s rights, and perform a separate assessment before commercial redistribution, dataset creation, or bulk archiving.

How can I reduce prompt-injection risk from captions?

Treat captions as untrusted quoted data, not instructions. Keep them out of the system prompt, tell the model to analyze rather than obey them, disable unrelated high-privilege tools, require confirmation for side effects, and audit every tool call produced from transcript-based workflows.

Bottom Line

YouTube Transcript MCP v0.7.0 is a focused community server with four read-oriented tools for existing captions, timestamps, metadata, and language discovery. Its current GitHub release, MIT license, Python 3.10+ baseline, expanded URL support, and .mcpb bundle make it practical for bounded research workflows. Its limits are equally important: it is not ASR, it cannot guarantee access to every video, PyPI is stale, and operators remain responsible for transcript rights, YouTube access, model-provider data transfer, caption accuracy, and prompt injection. Pin v0.7.0, check languages before retrieval, bound pagination and model spending, preserve source metadata, isolate the text from privileged tools, and verify conclusions against the original video.