Transcribe YouTube and podcasts from your AI agent with MCP

AI agents are good at reading. They are less good at listening, which is a shame, because a growing share of the content worth reading never gets written down: conference talks on YouTube, podcast interviews, voice notes. When I want an agent to summarise a talk today, the workflow is a clumsy relay: open a transcript tool in a browser, paste the link, copy the text, paste it into the chat, lose the tab.

The tautau MCP server cuts the relay down to one step. Ask the agent to transcribe the link, and the text appears in the conversation.

What it is

MCP (Model Context Protocol) is the standard way AI clients like Claude Code, Claude Desktop, and Codex pick up external tools. tautau-mcp is a small stdio server that gives those clients three tools:

• transcribe_url: paste a YouTube video, podcast episode, RSS feed, or direct audio link, get the full transcript back. YouTube resolves from the caption track in a second or two; podcasts and audio files run through AI transcription and can take a few minutes, which the server handles by polling for you.
• transcribe_audio: point at a local audio file (mp3, wav, m4a, ogg, webm, up to 25MB) and get it transcribed.
• get_quota: check how much of the free allowance remains.

It is the same transcription engine behind tautau.xyz, so the quotas and the privacy posture are the same: anonymous use is rate-limited per IP, and nothing is retained.

Install in thirty seconds

For Claude Code, one command:

claude mcp add tautau -- npx -y tautau-mcp

For everything at once, the setup script detects Claude Code, Claude Desktop, and Codex, and writes each config for you:

curl -fsSL https://tautau.xyz/mcp/install.sh | bash

Codex users who prefer to edit by hand add three lines to ~/.codex/config.toml:

[mcp_servers.tautau]
command = "npx"
args = ["-y", "tautau-mcp"]

Then ask for something useful: "transcribe this talk and pull out the three main arguments", or "turn this podcast episode into show notes with timestamps of the good bits". The agent calls the tool, waits through the processing state, and works with the text.

Where this gets interesting

A transcript inside an agent session is not a dead artefact, it is raw material. Some workflows that fell out naturally once the tool existed:

• Research digests. Feed an agent five talks from a conference playlist, ask for the themes they share and the claims they disagree on.
• Show notes that write themselves. A podcast URL in, structured notes with pull quotes out.
• Voice notes to tickets. Record a bug report on your phone, drop the m4a on the agent, get a formatted issue with reproduction steps.
• Meeting audio to action items. Transcribe the recording, then have the agent split decisions from loose ends.

The common thread: the transcription is the boring middle step, and boring middle steps are exactly what agents should absorb.

What it does not do

Worth being straight about the limits. YouTube videos without a caption track cannot be transcribed (tautau reads captions rather than extracting audio, which keeps it fast and keeps it on the right side of YouTube's rules). Long podcast episodes run against the async queue, so "a few minutes" is honest. And anonymous quotas are per IP, so a shared office network shares the allowance.

The server is open source, MIT licensed, and lives in the tautau repo under mcp/. It is also listed on npm, with Smithery and the MCP community registry submissions in flight. If your agent supports MCP, it is one npx command away from hearing.