Research API

A purpose-built research API for scientific agents: search papers, inspect metadata, read passages, expand the citation graph, and search research-related GitHub.

On the ArXivQA benchmark this stack reaches 61.0% recall — ahead of Firecrawl's Research Index (53.3%). It's live: results merge our own web and research search with live open scholarly sources, including full-text paper search — no self-hosted paper index.

Firecrawl's 53.3% is published in their own Research Index launch post (2026-06-17): "On arXivQA, the index hits 53.3% recall at $0.32 per task, against 45.4% for the next best provider." Our 61.0% is fastCRW's own measurement on the same 191-question ArXivQA set.

Authentication

Use a Bearer token: Authorization: Bearer <crw_live_…>. Get a key from the dashboard.

Endpoints

Both surfaces return identical shapes, served at the same /v1 path whether you call the cloud API or your own self-hosted engine. There are no research routes on the /firecrawl/v2 compatibility surface.

The sections below name this native /v1 path; the curl examples hit the cloud API (https://fastcrw.com/api).

Search papers

GET /v1/search/research/papers?query=&k=&authors=&categories=&from=&to=

curl -s -H "Authorization: Bearer $FASTCRW_API_KEY" \
  "https://fastcrw.com/api/v1/search/research/papers?query=diffusion%20image%20synthesis&k=20"

Returns ranked papers. paperId is the canonical id (a stable work id when known, else arxiv:<id>); primaryId is the preferred source id (arxiv:2105.05233); ids holds the prefix-less source ids.

{ "success": true, "results": [
  { "paperId": "W2105…", "primaryId": "arxiv:2105.05233",
    "ids": { "arxiv": ["2105.05233"] },
    "title": "…", "abstract": "…", "score": 0.42 }
] }

Filters: authors (substring), categories, from / to (YYYY-MM-DD).

Inspect a paper / read passages

GET /v1/search/research/papers/{id} returns metadata (authors, categories, createdDate, …). Accepts an arXiv id, a work id, or a DOI. Add ?query= to return the top passages answering a question:

curl -s -H "Authorization: Bearer $FASTCRW_API_KEY" \
  "https://fastcrw.com/api/v1/search/research/papers/arxiv:1706.03762?query=what%20is%20the%20attention%20mechanism&k=4"

Find related papers

GET /v1/search/research/papers/{id}/similar?intent=&mode=similar|citers|references&k=

intent is required. mode selects the expansion: similar (recommendations), citers (papers that cite the seed), references (papers the seed cites).

Search GitHub

GET /v1/search/research/github?query=&k= returns repository/README hits for implementation notes and engineering prior art.

SDK

The Firecrawl research SDK works unchanged — set the base URL to https://api.fastcrw.com. Or use the fastCRW SDK:

from crw import CrwClient
c = CrwClient(api_key="crw_live_…")
c.search_papers("diffusion image synthesis", k=20)
c.get_paper("arxiv:1706.03762", query="what is the attention mechanism")
c.related_papers("arxiv:1706.03762", intent="efficient transformers", mode="references")
c.search_github("flash attention implementation notes")

import { CrwClient } from "crw-sdk";
const c = new CrwClient({ apiKey: "crw_live_…" });
await c.research.searchPapers("diffusion image synthesis", { k: 20 });
await c.research.similarPapers("arxiv:1706.03762", { intent: "efficient transformers", mode: "references" });

The research skill

The endpoints are stateless primitives. The intelligence that reaches 61.0% on ArXivQA lives in the research skill over these endpoints, the same way Firecrawl splits its Research Index endpoints from its research skill. Its base pass is two steps, and both carry recall: 8-12 exact-name query reframings ranked by how many of them surfaced an id, and then mode=references on the top 5 ids to pull in the papers no query names directly. On top of that sit intent routing, reading a leaderboard, and pulling a paper's own bibliography for "compare-against" questions.

Measured on the first 12 benchmark questions with one raw query each, so the expansion is the only variable: 47.0% for the search union alone, 66.5% with the reference expansion. Skipping it is the most common way a re-implementation lands below the published number.

Install it into your agent (Claude Code, Cursor, Codex, Gemini CLI, …):

npx skills add us/crw@crw-research

Notes and limits

• Live, no index. Recall comes from merging our own web search (web + research mode) with live open scholarly sources + full-text paper search. Latency is seconds, not the milliseconds of a hot index.
• Read passages are abstract-scoped today; full arXiv-body passages are on the roadmap.
• GitHub results are repo/README-scoped today.
• Powered by open scholarly sources.