Undermind

Undermind takes a research goal in prose, asks follow-up questions, then runs what it calls a deep search: semantic search over the corpus, then citation trails, then the key authors, stopping when a pass adds nothing new. The vendor's own illustration of a run shows 340 papers evaluated in the first pass, 186 in the second and 94 in the third, then "Fourth pass added nothing new". A production search stops on its own after 2.9 minutes on average, and the FAQ puts a typical search at about two minutes.

Coverage comes from one source rather than many: "We cover all fields of science using the Semantic Scholar database of over 200 million articles, which contains PubMed, ArXiv, and many other databases." Full texts matter here, and they are tiered. The FAQ states that the Free tier searches abstracts and metadata, and that Pro adds compute to search full texts of nearly all open access and preprint articles. Paywalled PDFs are read only after a user uploads them to a workspace.

The output is not a link list. Reports are written into a workspace as Markdown with inline citations back to source papers, searches and starred papers persist, and alerts watch saved topics for new publications.

Undermind is published as a ChatGPT app and a Claude connector, with written setup for Cursor, VS Code, Codex, Claude Code, GitHub Copilot, Microsoft 365 Copilot via Copilot Studio, Opencode, Grok, Antigravity and Perplexity. The endpoint is https://mcp.undermind.ai/mcp over streamable HTTP.

The tool surface is unusually wide for a research service: deep search and fast search, author and metadata lookup, parallel full-text PDF reading, file writing, starring, and a full folder tree for curating papers. Display tools render cards in clients that support MCP Apps, and fall back to text with links elsewhere.

Authorization is OAuth with PKCE and a single mcp scope, and every request runs as a specific user inside that user's workspaces. There is no API key or service account: the reference says so plainly, which rules out a shared backend integration that is not acting for one person. Dynamic client registration is not supported, so an unlisted application either hosts a Client ID Metadata Document or emails support to be registered. Connections expire after 90 days without a refresh.

The pricing section carries a control most readers will miss: an Industry and Academic tab beside the monthly and annual switch. On the Academic list Pro is $20 a month, or $16 a month billed annually. On the Industry list the same plan is $75 a month, or $60 billed annually. That is a 3.75x difference for identical features, and no page states who qualifies as academic.

The discount is the vendor's own figure: "You save 20% with annual". Team is priced per person and quoted in blocks of five seats, $18.70 each monthly on the academic list and $75 on the industry list. Enterprise is custom priced and adds increased compute, sitewide organizational login, an admin dashboard and custom terms.

What a plan actually allows is never numbered. Free has "Standard rate limits on chats and searches" and Pro has "10x higher usage limits", a multiple of an undisclosed base, and the Acceptable Use Policy reserves limits on the number of searches or requests per user. Cancellation is self-serve at any time and access runs to the end of the paid period, but charges are non-refundable once a billing period has begun.

The enterprise material is unambiguous: "Your company's proprietary R&D data stays isolated, is never used to train AI models, and our model providers never retain or train on it."

For an individual account the Terms read differently. Section 3.3 grants the company a license to use User Content for providing the service, for "improving and developing the Services", and "in de-identified or aggregated form only, developing and improving related technologies, including artificial intelligence and machine learning systems, whether by us or our partners". Both statements can be true at once, because an enterprise agreement takes precedence where it differs, but a researcher reading the no-training card on the marketing page is reading a commitment that attaches to a contract they do not have.

User Content is defined broadly, and includes material retrieved on a user's behalf through integrations, agents and institutional access. The Terms put the rights check on the user: if you lack the rights to an item, do not upload it, and the company states it cannot verify authorization for any source.

The security overview lists encryption at rest and in transit, role-based access control with multi-factor authentication, daily backups with off-site storage, annual staff security training, and manual penetration testing by Oneleet with a certified tester. Evidence is gated: "Letter of Attestation / pentest report available under NDA."

What is absent is a named compliance regime. Neither the one-pager, the enterprise page nor the pricing table mentions SOC 2 or ISO 27001, and the Enterprise tier offers a "security review" rather than a report. Data location is single region: the privacy policy states that personal data is processed and stored in the United States, with Standard Contractual Clauses for transfers out of the EEA and UK. Payments run through Stripe. The company behind it is Undermind AI, Inc., registered in Dover, Delaware, founded by two MIT quantum physics PhDs, Joshua Ramette as CEO and Tom Hartke as CTO, and backed by Y Combinator.

The buyer this is shaped for is a researcher or an R&D team that needs recall rather than a quick answer, and will accept a few minutes per query to get it. GSK is the reference account, with 1,000 or more researchers using it inside an internal AI Scientist platform, and the enterprise logo wall adds Novartis, the NIH, Toyota Research and Fraunhofer IKTS.

It fits badly in three cases. A team that wants a stateless query API will find none: access is per user through MCP, and the "Undermind API" is an enterprise sales conversation with no public reference. Anyone whose literature sits behind paywalls does the fetching themselves, since only open access and preprint full texts are read automatically. And a buyer who needs a capacity commitment before signing has nothing to plan against, because no plan publishes a number for its limits.

The Terms also forbid building the product into another offering: the license is for personal or internal business use, and you may not "resell, repackage, or incorporate the Site or its functionality into any product, service, or offering".