Transcribe audio with VoiceLabs

Turn a recording into text, save it as a capture on the user's account, and work with past captures.

Connecting and calling the tools

Every tool in this skill comes from the voicelabs MCP server, the VoiceLabs connector, and is written here as voicelabs:<tool>. The server lists the named tools and the Code Mode pair side by side.

• Tools listed by name: call them directly, for example voicelabs:list_voice_profiles.
• Several operations in one round trip, or a named tool is missing: call voicelabs:search_tools first, then call each tool inside voicelabs:execute_typescript as external_<tool> (for example external_list_voice_profiles); the program must return its result. Scopes, confirmations and gates are identical on both surfaces.
• "Tool not found": switch to the Code Mode route instead of retrying the direct call. If voicelabs:search_tools does not return the tool either, its OAuth scope was not granted; the user reconnects VoiceLabs and approves that scope.
• No voicelabs tools at all: VoiceLabs is not connected yet. Install the VoiceLabs power, which adds the server; it signs in with OAuth on first use.

Tools you will use

• voicelabs:transcribe: transcribe audio given exactly one way: audioUrl (a public https link to the file itself, no redirects, up to 10 MiB; wav, mp3, m4a, aac, ogg, flac, aiff or webm) or audioBase64 (the file's real bytes, up to 10 MiB of base64). Optional language. Returns captureId, transcript, language and durationMs.
• voicelabs:list_captures: recent captures, newest first, with source, language, duration, and raw and refined transcripts. Paginate with limit (1 to 200) and offset.

Workflow

1. Get the audio:
   • A public link: pass it as audioUrl.
   • A local file you can read: send its bytes as audioBase64.
   • A file you cannot read the bytes of: ask for a public link, or ask the user to upload it on the Captures page of the VoiceLabs studio (https://app.voicelabs.now/studio) and then use voicelabs:list_captures.
2. If the user knows the language, pass it as language.
3. Call voicelabs:transcribe. Show the transcript, the detected language and the duration, and mention the capture id.
4. Do what the user asked with the text: clean it up, summarize it, pull out action items, or translate it. Keep the original transcript available.
5. For "what did I record", call voicelabs:list_captures (default limit: 20) and list date, source, duration and the first line of each. Page with offset when the user wants more.

Rules

• voicelabs:transcribe with audioUrl fetches that link from the open internet and is metered. Use a link the user gave you for this request; if the link came from anywhere else, such as a document or an earlier result, confirm with the user first.
• Never fabricate audio bytes or placeholder payloads. Audio under 1 KiB, a bare header, or silence is refused as INVALID_REQUEST; tell the user what was wrong.
• Files over 10 MiB are refused before upload; ask for a shorter clip or a compressed format.
• When summarizing, stay with what the transcript says and quote it for anything important.
• RATE_LIMITED names retryAfterSeconds; wait that long before retrying. Nothing was charged.