Skip to main content
Three tools cover finding and reading papers: search_literature is the core Scite MCP tool. The index covers 210M+ papers across all fields, including PubMed/MEDLINE. Requires a Scite premium subscription. Quoted text (excerpts, citation snippets, full text) returns only for openly licensed papers and papers your organization is licensed for.

What it returns

Every hit carries the paper’s identity plus the evidence attached to it. The response itself also carries:

Full-text status

Check fulltextStatus before claiming you saw a paper’s text.

Classifier confidence

Each citation’s confidence is the classifier’s 0–1 confidence in its type, absent when unclassified. Supporting and contradicting labels are conservative, so most statements land in mentioning. Treat a supporting or contradicting label at 0.8 or above as reliable, and weigh lower ones against the snippet text. A type of unclassified means the citation statement exists but the classifier hasn’t assigned it a category. It isn’t a fourth citation category — see Citation model.

Three ways to call it

Search for papers

Pass term with domain-specific vocabulary. The index spans every academic field, so broad phrases return noise.

Fetch metadata for known papers

Pass dois (preferred), pmids, or titles without a term. No search runs; you get the records back.

Read inside a specific paper

Pass dois with a term. Each call returns up to 5 matching excerpts, so vary the term across calls to walk through a paper section by section.
If fulltextExcerpts comes back empty, the full text either isn’t indexed or your terms didn’t match. Use the access field to reach the PDF instead.

Query syntax

term searches across title, abstract, and full text, and supports:
  • Boolean operators: AND, OR, NOT
  • Phrase search: "exact phrase"
  • Proximity: "term1 term2"~5

Parameters

Calling search_literature with no arguments browses the whole corpus, relevance-sorted. That’s allowed but rarely useful — pass at least a term, dois, or titles.

Query and pagination

Keep limit small. Every hit can carry excerpts and citation statements, so large limits produce payloads that consume a lot of model context. Use 10–50 with offset to page.

Metadata filters

Evidence filters

These filters use PubMed publication types and MeSH, so they narrow results to PubMed-indexed papers. Omit them to search the full corpus. Combine study_design, population, and mesh_terms for PICO-style searches.

Citation filters

Each filter is a range bound. tally counts citation statements; citingPublications counts distinct publications.
The API field is contrasting, matching the parameter names above. Scite’s interface labels the same category “contrasted.” Both refer to contradicting citation statements.

Editorial filters

Check editorialNotices on every paper before citing it, even when you haven’t filtered on these. A paper can be accurate and still be withdrawn.

Screening many candidates

For a systematic screen, keep each page small and drop papers you’ve already judged.
Then read the survivors with dois plus term, or with read_fulltext. Record the include and exclude decisions with report_citations.

Reading a paper efficiently

The token-efficient pattern is search once, then read incrementally.
1

Find the paper

Search with a narrow term and a small limit to get candidate DOIs.
2

Confirm it

Call again with dois and no term to pull metadata, tally, and editorialNotices.
3

Read sections

Call repeatedly with the same dois and different term values — methods, results, limitations — collecting 5 excerpts per call.
4

Follow the citations

Read the citations array for what other papers say about it, quoting snippet text as evidence. For more than the handful returned, page through them with get_citation_statements.

read_fulltext

read_fulltext returns a paper’s body text in order, sliced by character offset, so you can read the whole document instead of term-matched excerpts. It reads Scite’s own indexed text, not the publisher’s page or PDF.
To page, set offset to the previous offset plus returnedChars while hasMore is true. Offsets are stable within a session only, because re-indexing shifts positions.

Check the source

contentDenied is true whenever source isn’t fulltext, including when you get the abstract. The message field says why. In that case the response carries link, where the user can read the paper: their Article Galaxy linkout when they’re connected, otherwise the publisher page. agAccess carries the full access detail when available. get_pdf_link takes a doi and returns pdfUrl when an open-access copy is known, along with isOa, oaStatus, license, and link (the OA copy, else the publisher page). Use it when an agent needs to download, attach, or OCR a paper rather than read it in the conversation. For closed papers, read_fulltext still returns the abstract.