One long search, or a set of moves
Undermind bets that literature search is one careful, expensive operation. Give it a goal and it works through hundreds of candidates, then tells you how confident it is that nothing was missed. On a genuinely open question that is hard to beat.
Lune bets that an agent mid-task rarely has one clean question. It has a paragraph it is drafting and a number it half remembers. So the tools are small: one question, twenty-five angles, the citation graph, a typed table, a claim check. The agent picks the next call from the last result.
Coverage against selectivity
Undermind searches 200 million records. Lune reads what cleared review at a venue the field respects. On coverage this is not a contest.
It is a contest on what comes back. Everything Lune returns already passed peer review, so the model spends no judgment deciding whether a source counts. The cost is real: an arXiv-first subfield will have work we do not hold.
Both ship MCP, so look at the tools
Undermind's MCP server is thorough, free-tier, and works in every client worth naming. Anyone selling you agent access to the literature as a Lune-only capability is selling you something.
The shape differs. Much of Undermind's toolset manages its own workspace: folders, starred papers, saved searches. Lune's are evidence operations, and there is nothing to file afterwards because the output lands in your draft.
