VisionXLab/CitationClaw

What it solves

Researchers, grant writers, and evaluators often need to show how a paper has been cited and by whom, but digging through Google Scholar, metadata databases, and PDFs by hand is slow and error-prone. CitationClaw automates the whole process: you give it a paper title or a Google Scholar profile, and it crawls the citing papers, collects author and institution details, identifies renowned scholars, pulls the actual sentences where your work is cited, and packages everything into a standalone HTML report. That report can be reused in research summaries, grant applications, and academic presentations.

How it works

CitationClaw v2 runs a five-phase pipeline:

1. Citation entry discovery – finds the starting point for the paper or profile.
2. Citing-paper retrieval – uses Google Scholar through ScraperAPI to get the list of citing papers.
3. Author and metadata collection – pulls structured data from OpenAlex, Semantic Scholar, arXiv, Web of Science Starter API, and PDF-based fallbacks.
4. Scholar impact assessment – pre-filters, caches, and verifies renowned scholars using a search-capable LLM.
5. Citation-context extraction – downloads, parses, and reviews citing PDFs to recover the actual sentences that mention your work.
6. Report generation – creates Excel sheets, JSON, and a self-contained HTML dashboard with charts, a knowledge graph, citation descriptions, and cost summaries.

The architecture uses a Skills Runtime to register replaceable phase skills, and a TaskExecutor coordinates the flow. Model roles are separated: a search-capable LLM handles scholar verification, while cheaper lightweight models can handle report generation and citation-context extraction. There are three service tiers (Basic, Advanced, Full) to balance depth, runtime, and cost. For papers with more than 1000 citations, a year-traversal option splits Google Scholar queries by year to work around the search engine's display limit.

The v2 version is an architectural upgrade over v1: it relies on structured APIs first instead of pure LLM extraction, adds a PDF download/parse pipeline (v1 had none), and caches heavily to cut costs. The README reports that v2 reduces LLM token consumption to about 1/5 of v1 while improving the overall benchmark score from 42.33 to 68.98.

Who it's for

• Researchers who want to understand how their papers are being cited and by whom.
• Grant writers and evaluators who need concrete evidence of research impact.
• Academics preparing annual reviews, presentations, or tenure materials.
• Anyone who needs a shareable, self-contained report of citation impact without manual crawling.

Highlights

• PDF-grounded citation contexts – extracts the actual sentences that cite your work, not paraphrases; the largest improvement over v1 (+32.45 benchmark points).
• Structured metadata first – uses OpenAlex, Semantic Scholar, arXiv, Web of Science, and PDF fallbacks instead of unreliable LLM-only author lookup.
• Three service tiers – Basic, Advanced, and Full let you control depth, runtime, and cost.
• Cost-efficient – v2 reduces LLM token consumption to roughly 1/5 of v1.
• Self-contained HTML report – a single browser-readable file with charts, knowledge graph, cost summary, and an AI assistant entry point; it can be rebuilt from cache without re-crawling.
• Year traversal for large citation counts – handles papers with more than 1000 citations by querying Google Scholar year by year.
• Configurable integrations – optional ScraperAPI, Semantic Scholar API, Web of Science Starter API, MinerU for complex PDFs, and a CDP debug port for authenticated publisher downloads.
• Preflight model checks – the UI can test LLM connectivity before a full run.