Use this tool when a Deep Research or connector client needs citation-shaped NIH grant hits for a natural-language query. Returns {"results": [{id, title, text, url}]} where id is the NIH project_num to pass to fetch and url is the public OpenNIH grant page. Each record is PROJECT-level, not project-year: ids are unique within a response, and the record always describes the project's CANONICAL row — its latest fiscal year, resolved WITHOUT the query filter, so it is exactly the row fetch returns for that id. READ THE RETURNED TITLE BEFORE QUOTING A HIT: one NIH project number can span many sub-awards with unrelated titles (center grants such as P41/P50/P01 are the common case), so the returned title and snippet MAY CONTAIN NONE OF YOUR QUERY TERMS — the sibling row that matched your text is not necessarily the canonical row, and the canonical row is the page the citation opens. Do not assert that a returned grant is 'about' your query on the strength of the match alone; use search_grants when you need the row that actually matched. This is the OpenAI Deep Research compatibility wrapper around search_grants — prefer search_grants directly when you need filters (institution, IC, fiscal year, activity code), funding totals, or provenance. Bounds: query is REQUIRED and matched over grant titles at 2-64 characters (a blank query returns no results instead of the whole corpus); limit 1-50 (default 20) distinct grants per page; offset 0-10000 skips matching project-year rows exactly as search_grants does, so stepping offset by limit walks the full result set page by page — results past the first page are reachable, not invisible. A page is the de-duplicated projection of that row window: one project number owning rows in two windows (multi-component awards emit one row per sub-project) can appear on more than one page, and ids are unique within each response.