Search academic papers from Claude, Cursor and other AI assistants with MCP
Published September 26, 2026
AI assistants are good at explaining a topic and unreliable about where their claims come from. An MCP server for academic search closes that gap: instead of recalling a paper from memory, the assistant looks it up in a real index and cites what it finds.
What the OpenAbstracts MCP server provides
A single tool, search_papers, backed by an index of 10M+ papers from 100,000+ journals. Arguments:
| Argument | Meaning |
|---|---|
q | Search query (required) |
mode | papers (titles, abstracts, keywords, authors) or content (full PDF text) |
category | Subject filter, e.g. medicine, business |
yearFrom, yearTo | Inclusive publication years |
quartile | Journal quartiles, e.g. Q1,Q2 |
page, limit | Pagination (max 100 per page) |
Each result carries title, authors, year, journal, DOI, abstract, quartile and PDF/publisher links.
Setup
Add https://openabstracts.com/mcp as a custom connector in Claude and sign in on the Connect page — the step-by-step guide takes about a minute. Other clients can use the same URL with an x-api-key header.
Prompts that work well
- "Find peer-reviewed papers from the last five years on microplastics in drinking water and list the DOIs."
- "I claim that remote work increases productivity. Search for evidence for and against, and summarise what the abstracts say."
- "Check whether this reference exists: Obalade & Muzindutsi 2019, Adaptive Market Hypothesis and the day-of-the-week effect."
- "Search only Q1 medicine journals since 2022 for trials on intermittent fasting."
Keeping the assistant honest
Give the assistant a standing instruction, or install the citation Skill, so it always searches before citing, never invents a DOI, and says plainly when the index has nothing. The Skill is a short markdown file — read it, and adapt it to your own workflow.
What to expect
Results come from keyword and full-text matching over registered journals, so phrase queries the way a paper's title or abstract would. If the first query returns little, try synonyms or drop a filter. The index doesn't cover every paper ever published, so "not found" means "not found here".
Cost
MCP calls use the same billing as the REST API: 100 free searches a month, then $1.50 per 1,000, and empty searches are free.
Prefer code over connectors? See the curl quickstart.