start_paper_generation returns a generation_id straight away and the paper lands minutes later, so whatever drives it has to poll check_paper_status instead of waiting on the call; phase and percentage come back on every poll and the download URLs only appear at the end. Two things worth settling first: the token is copied out of a browser request rather than issued from a settings screen, and each run spends 1 to 3 credits depending on the page count you ask for.
An MCP front end to OpenPaper's research engine. You give it a topic; 18 specialized agents search 500M+ academic sources across OpenAlex, Crossref and Semantic Scholar and write a paper with every citation linked to a real publication. Three tools cover the whole surface: start a run, poll it, list what you have already generated.
- start_paper_generation takes a topic of up to 500 characters and returns a generation_id immediately rather than waiting for the paper
- check_paper_status reports status, progress percentage and the current phase against that ID, and hands back the download URLs once the run reaches completed
- list_my_papers returns up to 50 earlier papers with their topics, status and download links
- Output arrives as PDF, DOCX or ZIP, and citation style and length are part of what you ask for — APA 7th at 20 to 30 pages is a request the tool understands
Runs over stdio as uvx openpaper-mcp, which fetches the published package on first run, and reads OPENPAPER_API_TOKEN. There is no key page: you sign in at openpaper.dev, open the Network tab, and copy the value after Bearer from the Authorization header of any request. Papers are billed in credits by length — 1 credit for 5 to 30 pages, 2 for 30 to 50, 3 for 50 to 60 — and new accounts start with free credits.
One command plus a key — uvx openpaper-mcp, then supply credentials