Agents read Markdown far better than they read PDFs, slide decks or spreadsheets. Getting a document into that shape usually means leaving the conversation, opening a converter, downloading the result and pasting it back. The MCP server removes those four steps: the agent calls a tool and receives the converted file on disk.
What it exposes
Two tools, and deliberately only two.
list_conversionsanswers what a given format can become, read from the same table the website serves.convert_fileconverts one local file and returns the path it wrote.
Both run on the same credit balance as the website. A conversion started by an agent and one started in a browser cost the same and draw from the same account.
What it does not do
This list matters more than the one above, because the opposite claim is easy to make by accident. There is no batch mode, no webhooks, no job polling, no queue, no streaming, and no URL input. It converts one local file per call and waits for the answer. There is no public HTTP API and no command-line tool either: the server speaks stdio to a local agent, which is a different thing.
Where your key lives
The API key is read from the environment and from nowhere else. It is never a tool argument, so no prompt can talk a model into revealing it, and the server never writes it to a file. The plaintext key is shown once when you create it; only its hash is stored, so nobody can recover it for you, us included.
What it can convert
162 input formats and 3,649 conversions, each one proven by running it and reading the output back rather than by trusting a compatibility table. Documents, images, audio, video, ebooks, subtitles, archives, fonts, data files and bibliographies. Files up to 150 MB.