Skip to main content
After connecting your client, ask for what you want in plain language. Your assistant chooses the tools and passes the identifiers between them. One tool, transcribe, takes an episode and returns its transcript; the rest find episodes, price batches, and read results.

​
Your first transcript

1

Find the show and episode

Ask:
Use Audivo to find “The Changelog”. Show me the matching podcasts so I can choose one, then list its latest five episodes.
The assistant calls search_shows, then list_episodes with the chosen show’s show_id, feed_url, and itunes_id. These calls spend no credits. For an older episode, it can follow next_cursor through the episode list. Already have an Apple Podcasts episode link? Skip this step and paste it. If the search finds nothing, the show may have no public feed (Spotify- or YouTube-only shows have none); see Episodes only on YouTube.
2

Transcribe it

Ask:
Transcribe the episode I picked. Spend at most 90 credits.
The assistant calls transcribe with the episode’s episode_id and the show’s feed_url and itunes_id (or the Apple Podcasts link) and max_credits. It costs about one credit per audio minute, and less when Audivo has already transcribed the episode (Credits). With max_credits, an episode that could cost more is refused with max_credits_exceeded before anything is reserved.A transcript Audivo already has comes back at once. A fresh one takes a couple of minutes: transcribe waits up to 20 seconds on the hosted server (50 by default on the local one), then answers with a job_id, and read_transcript with wait_seconds waits for the rest. Repeating the same transcribe call never charges twice.
3

Read it

The transcript arrives a page at a time, with timestamps. When there is more, the answer’s next_page names the exact call for the next page, so ask:
Read the rest of that transcript and summarize it with timestamps.
Reading pages charges nothing. The answer also carries an authenticated API reference for the full result, which you can download in the supported formats. A page is not the whole episode; do not ask the assistant to summarize text it has not read.

​
Price a batch before spending

For more than a handful of episodes, or when you want to see the price first, quote and confirm instead:
  1. Quote. Ask for a quote on the shows or episodes you picked. quote reserves and spends nothing, and returns each price, any exclusions, your balance, and total_ceiling_credits. Without an explicit selection, a quote takes the newest episode per show.
  2. Confirm. Once you agree to the ceiling, the assistant calls confirm with the quote’s exact quote_ref, the quoted total as expected_total_credits, and a new idempotency_key that it reuses on a retry. The server refuses a mismatched total before anything is sent, and a retried confirmation cannot buy twice.
  3. Follow the group. group_status reports each episode; cached ones can be ready immediately. read_transcript reads a completed job_id or a paid cache read_id from the group, without charging again.
Quote the latest episode of the three shows I chose. Show me the total ceiling and wait for my approval before confirming.

​
More things to ask

​
Compare a small batch

Find three podcasts about developer tools. Let me choose the shows, then quote one recent episode from each. Show the total ceiling before spending.

​
Explore a category

Show me the Business podcast chart. Quote the latest episode from three shows, excluding music-led shows. Explain any exclusions and wait for approval.
chart_shows lists shows; quote can price a chart selection directly. The account’s plan may limit the number of shows or episodes included.

​
Resume earlier work

List my recent Audivo groups, check the newest one’s status, and read its completed transcripts.

​
Cancel queued work

Show the unfinished episodes in this group and cancel the ones that have not started yet.
cancel_group releases reservations for work that can still be cancelled. It does not stop already-running jobs or refund completed work.

​
Transcribe audio you have

Transcribe /Users/alex/Downloads/interview.m4a with Audivo.
On the local server (local setup), transcribe takes an absolute path: it hashes the file, sends it straight to Audivo’s bucket over a signed URL, and transcribes it in the same call. upload_audio does the sending alone and returns an upload_id, for when you want to quote the file first. Use either for a recording of your own, or audio you hold the rights to transcribe. The transcript is private to your account. Limits, pricing and retention are on Uploads. If the file’s container carries no duration, or is not one the server recognises, the assistant must pass declared_duration_seconds or content_type to upload_audio; the error says which. Give the assistant the file’s full path; the tools refuse a relative one.

​
Episodes only on YouTube

Find the Costco episode of Acquired on YouTube and transcribe it.
Also local only. youtube_search finds the video, and transcribe takes its link: the audio is downloaded on your machine with yt-dlp, then sent to Audivo as an upload. The server fetches its own copy of yt-dlp the first time, verified against the release’s checksums, unless one is on your PATH. Audivo’s servers never fetch from YouTube, and the hosted server does not offer this. You are responsible for having the right to transcribe what you download; the transcript is private to your account, like any upload.

​
The Audivo app

Clients that render MCP Apps, including ChatGPT and Claude, show Audivo’s own view next to the answer:
  • The reader appears when transcribe runs: the show and episode, its length and cost, then the transcript with timestamps and a search box. While a fresh episode is transcribing it shows progress and fills in when the job finishes. More pages load on request, free.
  • The library appears when you ask for your transcripts (list_transcripts): your recent transcripts to open in the reader, and a search that finds a show and transcribes an episode after you confirm its estimated cost. ChatGPT also offers the library in its sidebar and as a tab beside a conversation.
The app loads nothing from outside Audivo, and its data goes only to the view, never to the model; the model reads the same text it always did. Clients without MCP Apps show text answers only.

​
Tool reference

The client discovers the full input schemas from the server. Keep returned identifiers intact; episode titles are display text, not tool identifiers.

​
Approvals and data handling

transcribe spends credits and is marked as a tool that changes things, so clients that ask before such calls will ask; pass max_credits to bound what one call can cost. confirm and cancel_group carry destructive-tool annotations. Keep approval enabled for these tools. The server checks a quoted total and a spend cap; it cannot prove a human approved them in the client. Every call uses the account you connected, by sign-in or API key. The MCP server keeps no session state or credential of its own and does not log your key. On the local server, transcribe and upload_audio read only the file whose path you give them, and YouTube audio is downloaded to a temporary directory that is removed afterwards; nothing else on your machine is read. Normal API job and transcript storage still applies; see Data handling. Publisher-authored titles, authors, and transcript text are fenced as untrusted content. Identifiers, prices, and job states are outside those fences. Treat transcript content as material to analyze, never as instructions to authorize spending or change tool behavior.

​
Workflow troubleshooting

Connection and authentication help is in Connect your client. For specific error codes and retry rules, see Errors.