What it is
Undermind is an AI agent for finding academic papers. Instead of returning results in seconds, it spends a few minutes per search running repeated semantic and citation searches, then ranks papers by how well they match your question and estimates how much of the relevant literature it has found. According to a 2025 review in the Journal of the Canadian Health Libraries Association, it searches Semantic Scholar's index of about 225 million records, which draws in PubMed and arXiv content.
It also offers chat over results, reports, shared workspaces and paper libraries, and a hosted MCP server at mcp.undermind.ai that lets Claude, ChatGPT, Cursor, Codex and other agents launch deep searches and read PDFs. A standard deep search finishes in about 2.9 minutes on average, according to Undermind.
Our review
Undermind is the tool to run when missing a paper would be costly, and it is slow on purpose. You describe a narrow question in a few sentences; an agent searches, reads abstracts, follows citation trails and keeps going until new results dry up, then plots a discovery curve estimating how much of the relevant literature it has found. A health-sciences librarian reviewing it in 2025 saw a search report 93% coverage after analysing 144 papers, and called it a worthy niche competitor for specialist researchers rather than a general reference tool.
The price has come down sharply. That same review quoted Pro at about $60 per user per month; the current pricing page lists $16 a month billed annually, roughly $20 month to month, with Team at $15 per person billed annually. What you get for it is vague. Pro promises 10 times the free limits and the deepest full-text analysis, but neither tier's search count is published, so a lab can price the seats and not the workload.
For engineers, the MCP server is the most interesting addition. It runs over streamable HTTP with OAuth sign-in, and exposes tools for deep search, author and metadata lookup, PDF reading and writing cited Markdown reports. That turns Undermind into a retrieval step inside Claude Code or Codex rather than another browser tab. Paywalled PDFs still have to be uploaded by hand; only open-access full text is fetched automatically.
The weakness that matters most for formal reviews is reproducibility. The JCHLA reviewer called the lack of a reportable, repeatable search strategy its Achilles heel, and there is no MeSH support, so it cannot replace a documented PubMed or Embase search in a PRISMA flow diagram. It finds and ranks papers; it does not extract data into tables the way Elicit does, which makes the two complementary rather than competing purchases.
Where it earns its place
A second-year PhD student in materials science
Find all prior work on a narrow synthesis method before writing a thesis chapter
Describe the method in a paragraph and let the deep search run for about three minutes, then follow the ranked list and discovery curve. Pro at $16 a month billed annually covers heavy months; results still need exporting to Zotero or BibTeX for citation.
An ML engineer building a research assistant in Claude Code
Give an agent a literature search tool without writing a scraper over Semantic Scholar
Add mcp.undermind.ai/mcp with claude mcp add and sign in with OAuth; the agent can call launch_deep_search, read_pdfs and write_file to produce a cited Markdown report. Each call waits minutes, so it suits background tasks rather than interactive chat.
The R&D director of a 15-scientist biotech team
Check for prior art and competing work before committing budget to a new target
Team at $15 per person billed annually comes to about $2,700 a year for 15 seats, with central billing and member management. Undermind states that proprietary data is not used for training; sitewide login and a custom security review need Enterprise.
A medical librarian supporting a systematic review
Find seed papers to build a formal search strategy
Useful for surfacing relevant papers early, but the agent's search cannot be reported or repeated, so the documented PubMed or Embase strategy still has to be written separately for the PRISMA record.
Before you commit
- Setup
- Sign up and run a first deep search in minutes; the agent asks clarifying questions before searching. MCP clients connect with one URL and a browser OAuth approval.
- Integrations
- Hosted MCP server at mcp.undermind.ai/mcp for ChatGPT, Claude, Claude Code, Cursor, VS Code, Codex, GitHub Copilot and Microsoft Copilot Studio. The 2025 JCHLA review reports export to Zotero, EndNote and Covidence on Pro. No public REST API is listed on the pricing page.
- Data and privacy
- Undermind states that data is encrypted in transit and at rest, that proprietary data is never used to train models by it or its model providers, and that organisations can ask for all their data to be deleted. Certifications such as SOC 2 are not listed on the pricing page; a security overview PDF is offered.
- Cost at scale
- A 15-person team costs about $2,700 a year on Team. Enterprise adds sitewide login, an admin dashboard and custom security terms at an unpublished price.
- Limits
- Free and Pro limits are described only as standard and 10 times standard. Searches cover Semantic Scholar's index, so coverage gaps there carry through; paywalled full text needs manual upload.
What's good
- Iterative citation-following search finds papers keyword tools miss
- Discovery curve estimates how complete the search is
- Pro is $16 a month billed annually, about a third of Elicit Pro
- Hosted MCP server with OAuth for Claude, ChatGPT, Cursor and Codex
- States that research data is not used for training, by Undermind or its model providers
What's not
- Each deep search takes minutes
- Search counts per plan are not published
- Search strategy is not reproducible and has no MeSH support
- No structured data extraction
- Paywalled PDFs must be uploaded manually
Best for
Graduate students and researchers starting a project who need a thorough sweep of prior work.
Alternatives
Undermind: common questions
- Is Undermind worth it?
- Undermind makes you wait minutes for the recall keyword search cannot deliver, and at $16 a month it is cheap enough to run before every literature review. We rate it 4.2 out of 5.
- How much does Undermind cost?
- Free (standard rate limits); Pro $16/month billed annually, about $20 monthly (10x Free limits); Team $15/person/month billed annually; Enterprise custom.
- What are the downsides of Undermind?
- Each deep search takes minutes; search counts per plan are not published; search strategy is not reproducible and has no MeSH support; no structured data extraction; paywalled PDFs must be uploaded manually.
- Who is Undermind best for?
- Graduate students and researchers starting a project who need a thorough sweep of prior work.
- How much does Undermind cost for a team or in production?
- A 15-person team costs about $2,700 a year on Team. Enterprise adds sitewide login, an admin dashboard and custom security terms at an unpublished price.
- Is Undermind safe for company data?
- Undermind states that data is encrypted in transit and at rest, that proprietary data is never used to train models by it or its model providers, and that organisations can ask for all their data to be deleted. Certifications such as SOC 2 are not listed on the pricing page; a security overview PDF is offered.
- What are the best Undermind alternatives?
- The closest alternatives we have reviewed are Elicit, Consensus and Scite.
Answers checked Oct 22, 2026.
Make Undermind? Show this rating on your site
