Audivo

Audivo MCP server

[npm] [CI]

Podcast transcripts for Claude, ChatGPT, Codex, Cursor, and any other MCP client, backed by the Audivo API. Ask for an episode; get the transcript back.

Transcribe the latest episode of Acquired and summarise it.

Two ways to connect. The hosted endpoint serves eleven tools. The local server serves the same eleven, with transcribe also taking a file on your machine or a YouTube link, plus upload_audio and youtube_search. Both serve an app to clients that render MCP Apps.

Tools

transcribe

Pass one episode: an Apple Podcasts link, feed_url with guid, an episode_id from list_episodes (with that show's feed_url and itunes_id, so an episode nobody has transcribed yet can be found), or an upload_id — and on the local server, a YouTube link or an absolute path.

• An episode that is already transcribed comes back at once.
• A fresh one becomes a job. transcribe waits for it inside the call: up to 20 seconds on the hosted server, whose gateway allows 29, and 50 by default on the local one, which reports progress to clients that ask for it. If it is still running, the answer is the job_id and read_transcript waits the rest.
• The transcript comes a page at a time, about 40,000 characters each (roughly 50 minutes of speech). Each page names the exact call for the next one.
• max_credits refuses the call, before anything is spent, if it could cost more. A job holds its ceiling (the estimate plus 25%) and settles at the audio it measured, never above.
• The same call twice returns the same job, never a second charge: the idempotency key is derived from the request.

For many episodes at once — a chart, a back catalogue — use quote and then confirm, which refuses unless the model restates the quote's total.

The app

Clients that render MCP Apps — ChatGPT, Claude, and others — show results in Audivo's own view instead of only text:

• transcribe opens a transcript reader: the show and episode, its length, language and cost, then the transcript with timestamps, a search box, and more pages on demand. A job still running shows its progress and fills in when it finishes.
• list_transcripts opens your library: recent transcripts to open in the reader, and a search that finds a show, lists its episodes, and transcribes one after you confirm its estimated cost. ChatGPT also puts the library in its sidebar and as a tab beside a conversation.

The app is one self-contained page served as ui://audivo/app-v1.html. It loads nothing from any domain, so it declares none. Its data travels in each result's _meta, which hosts give the app and never the model; the model keeps getting the same fenced text as before. Every title and transcript line is rendered as text, never as markup. Clients without MCP Apps ignore all of this.

Its source is in ui/; npm run build:ui bundles it into src/app-html.ts, and a test fails when the two disagree.

Local: run it with npx

You need an API key from the Audivo dashboard. Keys are shown once. Set the key in the environment the client starts the server with, then add the server.

Claude Code (or install the Audivo plugin, which adds this server and the skill together)

claude mcp add --scope user audivo -e AUDIVO_API_KEY=hk_live_... -- npx -y @audivo/mcp@latest

Codex

codex mcp add audivo --env AUDIVO_API_KEY=hk_live_... -- npx -y @audivo/mcp@latest

Cursor, Claude Desktop, Windsurf, and other JSON-configured clients

{  "mcpServers": {    "audivo": {      "command": "npx",      "args": ["-y", "@audivo/mcp@latest"],      "env": { "AUDIVO_API_KEY": "hk_live_..." }    }  }}

VS Code (.vscode/mcp.json)

{  "servers": {    "audivo": {      "type": "stdio",      "command": "npx",      "args": ["-y", "@audivo/mcp@latest"],      "env": { "AUDIVO_API_KEY": "hk_live_..." }    }  }}

Keep files that contain a real key out of Git and shared chats.

Updating

@latest makes npx look the package up on npm each time the client starts the server, so a new release arrives with your next session. To get it in the session you are in, restart the server: in Claude Code, /mcp, choose audivo, then Reconnect. A spec without @latest works too, but a bare @audivo/mcp resolves to a copy in the current project first, when there is one. The hosted server at https://api.audivo.dev/mcp is kept current by Audivo.

Environment

The server writes nothing to stdout except protocol messages. Log lines go to stderr as JSON, and the key never appears in them.

YouTube (local server only)

When a show has no public RSS feed — many exist only on YouTube — search_shows finds nothing. On the local server, youtube_search finds the episode and transcribe takes its link:

Find the Costco episode of Acquired on YouTube and transcribe it

transcribe downloads the video's audio on your machine with yt-dlp, audio only and with no transcoding, uploads it to Audivo as your own upload, and transcribes it. The transcript is private to your account.

yt-dlp is found in this order: AUDIVO_YTDLP_PATH; a yt-dlp on your PATH; otherwise the project's official standalone binary for your platform, downloaded once from its GitHub releases into your user cache directory and checked against that release's SHA2-256SUMS before it is made executable. The package has no install script: nothing is downloaded until the first YouTube call. A managed copy that stops working is updated to the latest release once and the call retried; a yt-dlp you installed yourself is never touched. yt-dlp is given the Node running this server as its JavaScript runtime.

You run the download, on your machine, under your own account; you are responsible for having the rights to transcribe what you download, as with any upload. Audivo's servers never fetch from YouTube, which is why the hosted server does not offer any of this.

Upload your own audio

A recording you made, an interview, or any file you already have on disk:

Transcribe /Users/alex/Downloads/interview.m4a

transcribe with path announces the file to Audivo (its hash, size, content type, and duration), uploads the bytes straight to Audivo's storage with the signed URL the announcement returns, and transcribes it. upload_audio does only the first two steps and returns an upload_id, for when you want to quote several uploads together. The path must be absolute.

Limits: 1 byte to 5 GiB, up to 10 hours, and one of these content types: audio/mpeg, audio/mp3, audio/mp4, audio/m4a, audio/x-m4a, audio/aac, audio/x-aac, audio/ogg, audio/opus, audio/flac, audio/x-flac, audio/wav, audio/x-wav, audio/webm. An upload is kept for 7 days, and its transcript is private to your account; each account may hold up to 10 GiB across 100 unexpired uploads at a time. See the uploads guide.

Hosted: point a client at the URL

Clients that support MCP authorization — Claude, ChatGPT, Claude Code, VS Code — need only the URL: they open an Audivo sign-in page, you approve the connection, and it appears under Connected apps in the dashboard, where you can revoke it. For example, in Claude Code:

claude mcp add --transport http --scope user audivo https://api.audivo.dev/mcp

For clients that take a fixed header instead, or for automation, send an API key:

claude mcp add --transport http --scope user audivo https://api.audivo.dev/mcp \  --header "Authorization: Bearer hk_live_..."

Per-client instructions are in the connection guide. The hosted server is this package's lambda export, deployed by Audivo.

How it works

• Stateless. Every tool call is one or a few typed calls on the public API with the caller's own credential. The server keeps no table, no cache, and no copy of a credential beyond the call in flight.
• Fenced. Show names, episode and video titles, descriptions, and transcript text are publisher-authored. Each reaches the model inside a fence marked as untrusted content, with a per-response nonce, so a podcast cannot smuggle instructions into a session.
• Bounded. transcribe spends within the job's ceiling and the caller's max_credits; confirm compares the total the model states with the total the quote carried and refuses on a mismatch without sending anything. The API applies both checks on its side too.
• Typed by the contract. src/contract/types.ts is generated from the published OpenAPI spec in contract/openapi.yaml; a test fails the build when the two drift.
• Local only. upload_audio, youtube_search, and the file and YouTube inputs of transcribe run on your machine; the hosted server never sees your files and never fetches from third-party sites.

Development

npm cinpm test            # vitestnpm run typechecknpm run lintnpm run build       # dist/

To pick up a spec change: npm run contract:sync fetches the published spec and regenerates the types. Run it locally against a key with:

AUDIVO_API_KEY=hk_live_... node dist/bin.js

Live reload in your MCP client

npm run dev:server runs src/ through tsx behind a small proxy that restarts the server on every save, replays the client's handshake into the new process, and sends tools/list_changed. A connected client uses your latest code with no build and no reconnect. A save that breaks startup answers calls with an error, the details go to the client's MCP log, and the next save recovers. A call still running when you save is sent again to the new code, except one that uploads audio (a file or YouTube transcribe, or upload_audio): sent twice, that could be transcribed and charged twice, so it finishes on the old code first.

To use it in Claude Code for this repository only, taking precedence over a user-wide audivo server:

claude mcp add audivo -s local -e AUDIVO_API_KEY=hk_live_... \  -- "$PWD/node_modules/.bin/tsx" "$PWD/scripts/dev-server.ts"

The absolute tsx path lets the client start it from any directory. A bare npx -y @audivo/mcp does not work inside this repository — npm takes the checkout to be the package and finds no linked audivo-mcp bin — which is one reason the configs above say @latest.

Releasing

Bump version in package.json, both version fields in server.json, and SERVER_INFO in src/server.ts (a test fails until they agree), add a CHANGELOG.md entry, commit, then tag v<version> and push the tag. The release workflow publishes to npm with provenance and lists the release in the MCP Registry as io.github.AudivoDotDev/mcp.

License

MIT