Literature search agent that clarifies your question, follows citation trails across 200M+ papers, and links each statement to its source paper.
Researchers doing a hard, specific search where the goal is to find every relevant paper rather than get a quick summary: the agent asks follow-up questions, evaluates papers and follows citation chains before stopping on its own, which the vendor's benchmark puts at about 2.9 minutes per search. It suits people who then want to work on those papers, with custom tables, full-text reading, alerts for new publications and in-line citations that link back to source papers. The free plan covers deep searches and reports, so the claim can be tested before paying.
Do not treat it as a full-text engine for subscription literature: the MCP page states open-access PDFs are fetched automatically but paywalled PDFs can only be read after you upload them to the workspace yourself. Buyers who need to audit corpus coverage will be stuck, because the site states 200 million+ papers without ever naming the databases behind that figure, and the FAQ entry on data sources renders without an answer. There is no documented REST API, Pro is only priced as $16/month billed annually with no monthly figure shown, and the free plan's limits are described as 'standard rate limits' with no numbers.
Pros
Cons
The MCP page states semantic search over 200 million+ papers, and the homepage says a search evaluates hundreds of papers and follows citation trails. The databases behind that number are not named anywhere on the site, and the FAQ question about data sources appears without an answer. If exact coverage of a field matters to you, confirm it with Undermind before relying on a negative result.
Yes, which is unusual. At the ten-minute mark it reports 85% recall of the twenty most relevant papers versus 50%, 47% and 43% for three general chat models, 75% recall of all relevant papers versus 38%, 33% and 30%, and 58% precision versus 40%, 35% and 40%. The caveat is in the method: 23 queries chosen by Undermind, graded by an LLM judge over abstracts.
Not on its own. The MCP page states open-access PDFs are fetched automatically, while paywalled PDFs can be read after you upload them to the workspace. Pro is advertised as the deepest analysis of full texts, but that does not change access rights. Institutional subscribers should expect to supply PDFs manually for anything behind a publisher paywall.
No REST API is documented. Programmatic access is through an MCP endpoint at mcp.undermind.ai/mcp, with published setup steps for ChatGPT, Claude web, desktop and Code, Cursor, VS Code, Codex, GitHub Copilot, Microsoft Copilot, Opencode, Grok, Antigravity and Perplexity. Rich result cards render only in clients supporting MCP Apps, such as ChatGPT and Claude; other clients receive the papers as text with links.
Facts sourced from the vendor’s own site, last verified 2026-10-10.
The Weekly Source
One email a week. New listings, repriced tools, and the ones we would skip - including this one if its pricing changes.
