pdftomp3 turns PDFs into audiobooks using your AI chat subscription

A hosted MCP server charges nothing of its own because the compute is already in the user's AI chat subscription. pdftomp3, released this week by Jan Metz as a Show HN that reached 169 points and 90 comments, turns a PDF into an audiobook from inside Codex or Claude. The user pays zero to pdftomp3 and pays OpenAI or Google only for the final speech synthesis.

The connect page states the split plainly. Under "Your subscription will be used for" sits "document processing (via MCP)" with the note "already paid by your subscription", meaning the AI chat agent is the one that spends the tokens that read the PDF. Under "Your API key will be used for" sits "generating speech", with the note that OpenAI or Gemini "charges per usage". The headline cost line reads "$0 to PDF TO MP3". No paid plan from pdftomp3 appears at any point in the signup flow.

Where the money actually goes

Two external bills are visible, neither going to pdftomp3. The AI chat agent charges the user's existing AI subscription for the tokens it spends parsing the PDF and shaping the request. The speech provider charges the user's own key for every audio minute. pdftomp3 sits between them, handles the file parsing and the output assembly, and takes no cut. The page notes that a free Gemini key can hit rate limits, so a billing-enabled Google key is more reliable.

That is a smaller job for pdftomp3 than it looks. The chat tool handles the model work. The speech API handles the voice. pdftomp3 handles the file and the orchestration. The server's job is thin enough that whatever hosting cost it carries is covered by other revenue, including the paid plans pdftomp3 sells on its pricing page for users who prefer not to connect their own AI account.

Why this shape is worth watching

A hosted MCP server lets the user's existing AI subscription cover the heavy call, which is the call that would have forced the vendor to run its own paid plan. The vendor can be free because the vendor is not paying for the model that reads the text. The pattern is older than this product; what pdftomp3 shows is how little the server has to do once the agent is doing the reading.

For a team shipping a thin agent-native product, the shape is practical: a hosted MCP server, one or two external API keys the user supplies for the parts that still cost money, and a product small enough to run cheaply. Jan Metz posts from @janxpm on X and Bluesky; the full product sits at pdftomp3.com.

Source

• Primary: pdftomp3 connect page
• Show HN discussion, 169 points