retrievr

For agents

An MCP server your agent can ask about PDFs

Upload a PDF once. From then on your agent asks a question in plain language and gets back the passages that answer it — each one naming the page it came from, so the answer can be checked instead of trusted.

What comes back

Not the file. A short list of passages, usually two or three, packed to about 400 tokens each. A passage is a window on one page and never spans two, so a citation cannot quote the end of one page as though it sat beside the start of the next.

Every passage carries its source, author, section, character range and licence. That is what makes an answer checkable: the agent quotes what it retrieved instead of reconstructing it, and a wrong answer can be traced to the search that missed or the reading of what it found.

A page whose text is drawn as an image has no text to return, and a PDF made only of those is refused at upload with a message saying so. There is no OCR here — a scanned document is reported as unreadable rather than accepted and quietly answering nothing.

Give your agent the tool

Paste this into the agent you already use, and let it work:

Fetch https://retrievr.io/install and follow it.

The sheet gets a key without printing it, downloads the MCP server, and shows the entry for your harness — Claude Code, Codex, Cursor, pi, opencode, Kiro or Hermes. The documentation carries each harness’s file edit, and the Claude Code guide walks the same flow by hand.

Five tools come with it:

retrievr_upload puts files in your library from paths on your machine, so their contents never pass through the conversation. retrievr_search takes the question. retrievr_documents lists what is there. retrievr_upload_status says when a file is ready, or which one failed and why. retrievr_delete_document removes one for good.

PDFs, and three other formats

PDF, EPUB, plain text and Markdown, up to 50 MB each and twenty files per call. Each format is read the way it is built: a PDF page by page, an EPUB in spine order with the table of contents and boilerplate dropped.

An EPUB chapter becomes a section the same way a PDF page does, so a citation from a book names the chapter and one from a report names the page.

What it does not do

It does not convert a PDF to Markdown, and it does not lay a page out for you. If the job is reformatting a document, a converter is the right tool.

It is not a chat window over your files. There is no interface to read — the caller is an agent, and the answer is passages it can cite. If you want to read and search a PDF yourself, a document reader will serve you better.

There is nothing to operate. No vector database to run, no embedding job, no index to keep in step — uploading a file is the whole setup. The service is always ready: a query is answered in well under a second.

What it costs

$0.01 per query, charged when the query is accepted. Uploading and indexing are free: the first five documents need no payment at all, and after that one $5 top-up unlocks the rest and is spent on queries. The full price list.

Questions worth asking before you start

Does my PDF leave my machine? A local install reads the file from the path you name and sends it to your own library; the contents are never pasted into the conversation, so they are not in the transcript.The privacy page is specific about what is kept and for how long.

Can it search one document? Yes — name it and the search covers only that. Otherwise a search covers your whole library, and there is no public corpus behind it to widen into.

What if a question matches nothing? You get an empty result, and an empty result is an answer: the document either is not in the library or does not say what was asked.