Skip to main content
Firecrawl Research is a purpose-built index for scientific and engineering research agents. It exposes a research-specific toolset for searching papers, inspecting paper metadata, reading relevant full-text passages, and discovering related papers through structural expansion. The index covers roughly 43 million paper abstracts. The majority of the corpus is biomedical and life sciences — PubMed, bioRxiv, and medRxiv — alongside arXiv for physics, mathematics, and computer science. Papers are addressable by their source ids, so pmid:, pmcid:, and doi: references work the same way arxiv: ones do.
  • Find papers by topic, method, benchmark, author, or category
  • Inspect canonical paper metadata and source ids
  • Read the passages in one paper that answer a specific question
  • Expand from strong seed papers to related papers, citers, or references
  • Look up a protein’s curated location and function, with the evidence kind for each statement and the papers UniProt cites
To give your agent access to the Research Index, we strongly recommend using our CLI or MCP, combined with our dedicated research skill, which you can install with:

How this relates to /search

Firecrawl has two things named “research”, and they are not the same feature: Use the Research Index when you are doing literature work: finding papers, reading them, and following citations. Use categories: ["research"] when you want ordinary web pages that happen to live on academic domains.

Endpoints

The research index GitHub search was removed. GET /search/research/github and the legacy GET /research/github were removed on 2026-11-03. Both paths now return 410 Gone. Search GitHub issues, pull requests, and READMEs with the Developer Index instead.
Research Index calls are free. They have a limit of 100 a minute on every plan and a daily limit set by your plan, and Scale and Enterprise plans have no daily limit. Searches, reads, inspects, related-paper lookups and protein lookups all count toward both, including calls to the Research Index tool in Alexandria. See rate limits for the limit on each plan.

Search papers

Search paper abstracts with a natural-language query. The response returns ranked papers with canonical paperId, preferred primaryId, source ids, title, abstract, authors, creation and update dates, score, and optional ranking signals.
Optional filters:
  • authors: author substring filter; all filters must match
  • categories: paper category filter, such as cs.LG
  • from: inclusive created/updated lower bound, YYYY-MM-DD
  • to: inclusive created/updated upper bound, YYYY-MM-DD
  • sort: relevance (default) or date_desc, which returns the newest of the most relevant matches, newest first by created date. It does not list every paper in a date range; add from (and to) for that.

Biomedical example

Most of the index is life sciences, so clinical and molecular biology queries work the same way:
Results carry a primaryId in whichever namespace the source uses, so biomedical hits come back as pmid:<id>, pmcid:<id>, or doi:<doi>. Feed that value straight back into the inspect, read, and related endpoints below.

Inspect a paper

Use a canonical paperId or a source-specific primaryId. Accepted primaryId forms are arxiv:<id>, pmid:<id>, pmcid:<id>, and doi:<doi> — for example curl -s ".../v2/search/research/papers/pmid:<id>" for a PubMed record.

Read paper passages

Add query to the same paper path to retrieve the top full-text passages for a question. This is useful for verifying whether a candidate paper actually contains a method, dataset, constraint, or result before you include it.
Expand from one or more seed papers through semantic expansion and rank the candidates against a natural-language intent.
Modes:
  • similar: co-citation and bibliographic-coupling neighborhood
  • citers: papers that cite the seed
  • references: papers cited by the seed

Look up a protein

Look up one named protein and get what UniProtKB/Swiss-Prot curators record about it: its subcellular location, its function, and its GO cellular-component terms, plus its disease, PTM, subunit, tissue specificity and domain annotation, its sequence features and its interactions. Each statement carries its evidence kind and, where UniProt cites one, the PubMed paper behind it as a pmid: id, in the form inspect and read accept. Swiss-Prot is the reviewed section of UniProtKB, and an entry outside it returns not_found.
Parameters:
  • protein: one identifier. A UniProt accession (Q6P087), an entry name (RUSD3_HUMAN), a gene symbol (RPUSD3), or an exact protein name (Mitochondrial mRNA pseudouridine synthase RPUSD3).
  • organism: the organism a gene symbol or protein name is looked up in. An NCBI taxon id (9606), a scientific name (Homo sapiens), or a UniProt common name (Human). Defaults to 9606, human.
The response below is shortened. The full one has two more GO terms, the whole function text, and the fields described after it.
subcellularLocation and function hold UniProt’s statements in UniProt’s order. A location statement with note: true is the free-text note of a location comment. goCellularComponent lists only the GO terms with one of these experimental GO evidence codes: IDA, HDA, IPI, IMP, IEP, EXP, HTP, IGI. A GO term’s paperIds is empty when no paper is loaded for it, and only GO terms of human entries have paper ids. An entry also carries these fields, which the example leaves out:
  • disease, ptm, subunit, tissueSpecificity and domain: UniProt’s statements on those topics, in UniProt’s order, with the same text and evidence as a function statement. A disease statement names its disease in diseaseName, with diseaseAcronym and the OMIM id in mimId when UniProt gives them.
  • features: up to 50 sequence features in UniProt’s order, such as chains, domains, modified residues and binding sites. Each has a type, start and end residue positions, a description, UniProt’s feature id, a ligand and ligandId on binding sites, and evidence. featureCount gives the full number. Natural variants, mutagenesis results and isoform sequences are left out, and variantCount gives their number.
  • interactions: up to 25 binary interactions from IntAct, those with the most experiments first. Each names the partner accession and partnerGene, says whether the partner is from another organism in organismDiffers, and gives the number of experiments and the intactIds. Interactions carry no evidence kind and no paper. interactionCount gives the full number.
  • length and mass: residues and daltons of the canonical sequence.
  • pdbCrossReferenceCount: how many PDB structures the UniProt entry cross-references.
The entry’s paperIds collects every paper cited anywhere in the entry once, so it can include papers cited only in disease, ptm, subunit, tissueSpecificity, domain or features. To find the statement a paper supports, look for its paperId in the evidence items. From MCP, call firecrawl_research_lookup_protein with protein and, optionally, organism.

How a lookup resolves

The lookup matches whole identifiers and does not read free text, so a question such as “where is RPUSD3 localized” returns match: "not_found". Use paper search for questions. Accessions, entry names, gene symbols and protein names match in any case.
  • A primary accession or an entry name names one entry. An isoform suffix such as -2 is dropped, and organism does not narrow the match.
  • A secondary accession is an older accession that UniProt keeps on the current entry, and a match on one has matchedBy: "secondary_accession". When UniProt split an entry, its old accession stays on each new entry, and the lookup returns them as ambiguous.
  • A gene symbol or a protein name is looked up in organism. A primary gene symbol wins over a synonym. A protein name is a recommended or alternative full name.
  • When the symbol or name fits one entry, match is found. When it fits several, match is ambiguous. candidates lists up to 10 of them, with identifiers and no annotation, and candidateCount gives the full number. Call again with one candidate’s accession.
  • When nothing in organism matches, the lookup checks other organisms and returns what it finds as ambiguous, even when that is a single entry. The lookup never guesses between proteins and never switches species silently.
  • Anything else is not_found, with status 200.
A protein that is missing, empty, or longer than 200 characters, an empty organism, or an unknown query parameter returns 400 with error: "Invalid query parameters" and details. An organism that matches no organism in the data, or a name shared by more than one taxon, returns 400 with code: "BAD_REQUEST" and an error that says what to pass. The organism check runs before the lookup, so it applies even when protein is an accession. A lookup that runs past its 2-second limit returns 504 with code: "GATEWAY_TIMEOUT". If no answer arrives within 5 seconds, the API returns 504 with an empty body.

Evidence kinds

Each statement has one or more evidence items. kind names the type of evidence, and eco gives the Evidence and Conclusion Ontology (ECO) code that UniProt attaches. See UniProt’s evidence guide for the full definitions. An evidence item has a paperId when UniProt cites a PubMed paper for it. A by_similarity or sequence_analysis statement is inferred, not measured on this protein.

Attribution

Protein data from UniProtKB/Swiss-Prot. Copyright UniProt Consortium. Licensed under CC BY 4.0. No warranty, see the UniProt license. Selected fields, reformatted. Every 200 response repeats this in notice and names the UniProt release in release.