Claw4Science

Automate Your Lit Review With AI Agents

Apr 3, 2026

Key Takeaways

  • 10+ literature search skills exist across major skill hubs — but quality varies wildly
  • Best all-rounder: openalex-database — 240M papers, zero API key, all fields
  • Best for biomedicine: pubmed-database — MeSH queries, publication type filters
  • Best for systematic reviews: deep-research (Imbad0202) — 13 agents, PRISMA-compliant
  • Best for finding gaps: lit-synthesizer (ClawBio) — citation graph + gap analysis
  • You can chain these into a full pipeline: discover → filter → synthesize → cite

The Literature Review Problem

Every researcher knows the drill. You start a new project. You need to know what's already been done. So you open PubMed, type some keywords, and start reading.

Three weeks later, you have 47 browser tabs open, a spreadsheet with 200 papers you haven't actually read, and a growing suspicion that you've missed something important buried on page 8 of the search results.

Literature reviews are the tax every scientist pays. They're essential — you can't do good research without knowing the landscape. But they're also brutally time-consuming, repetitive, and error-prone. Miss one key paper, and your reviewer will find it.

AI agents don't eliminate this work. But they can compress three weeks into three hours.

The Tools: What's Available

We surveyed every science-focused skill hub on Claw4Science and found 306 skills related to literature search across 12 repositories. That's the most crowded category in the entire ecosystem — more than genomics (550), more than clinical tools (216).

But most of those 306 skills are minor variations of the same thing: a PubMed wrapper. The genuinely useful ones fit on one hand.

The Starting Five

1. openalex-database (Editor's Pick)

What it searches: 240 million scholarly works across all fields API key needed: No Install: clawhub install openalex-database

This is where you start. OpenAlex is the largest open academic database in the world — every field, every journal, completely free. No API key, no rate limits worth worrying about, no paywall.

It does things PubMed can't: author network analysis, institution tracking, citation counting, open-access discovery. If you need to answer "who are the top 10 labs working on X?" or "how has the citation count for this topic changed over the past 5 years?" — this is your tool.

When to use it: First pass on any topic. Broad landscape mapping. Citation analysis. Finding open-access versions of paywalled papers.

2. pubmed-database

What it searches: PubMed (biomedical and life sciences) API key needed: No (optional for higher rate limits) Install: clawhub install pubmed-database

If you're in biomedicine, this is your workhorse. What makes it better than just typing into pubmed.gov? MeSH term support. You can filter by publication type — only randomized controlled trials, only meta-analyses, only systematic reviews. You can build complex Boolean queries that would take 20 minutes to construct manually.

The skill handles the NCBI E-utilities API properly — batching requests, respecting rate limits, parsing XML responses into clean structured data.

When to use it: Biomedical research. Systematic reviews that need MeSH-controlled vocabulary. Finding clinical evidence by study type.

3. deep-research (Imbad0202)

What it does: Full autonomous literature review pipeline API key needed: Depends on mode Install: git clone https://github.com/Imbad0202/academic-research-skills.git

This isn't a search skill — it's a research machine. Thirteen AI agents work in sequence across seven modes: full research, quick brief, paper review, PRISMA-compliant systematic review with meta-analysis, Socratic guided dialogue, and fact-checking.

The PRISMA mode is remarkable. It designs a search strategy, executes it across multiple databases, screens results by title and abstract, applies inclusion/exclusion criteria, assesses risk of bias, and optionally runs a meta-analysis. It outputs an APA 7.0 formatted report.

The catch: It consumes enormous token budgets — 200K+ input and 100K+ output for a full run. But for a genuine systematic review, that's still cheaper and faster than doing it manually.

When to use it: Formal systematic reviews. Grant proposal background sections. Comprehensive literature surveys where thoroughness matters more than speed.

4. lit-synthesizer (ClawBio)

What it does: Literature synthesis with citation mapping and gap detection API key needed: No Install: clawhub install lit-synthesizer

Most search tools find papers. This one finds what's missing. It searches PubMed (with MeSH terms) and bioRxiv/medRxiv, then builds a citation relationship graph and runs gap analysis — identifying areas where keyword coverage drops off, suggesting under-studied intersections.

Imagine searching for "CRISPR + liver cancer" and getting back not just the papers that exist, but a map showing that "CRISPR delivery methods specific to hepatocytes" has been studied extensively while "CRISPR + liver cancer + immunotherapy combinations" has almost no coverage. That gap is your next paper.

When to use it: Finding research gaps. Identifying under-studied topic intersections. Building literature review sections that go beyond listing papers to analyzing the field's structure.

5. citation-management

What it does: Multi-source citation search + BibTeX generation + anti-hallucination API key needed: No Install: clawhub install citation-management

The cleanup tool. After your AI agent has generated a draft with 50 citations, how do you know they're all real? LLMs hallucinate references constantly — inventing plausible-sounding author names, journal titles, and DOIs that don't exist.

This skill searches Google Scholar, PubMed, CrossRef, and arXiv to verify every citation. It converts DOIs, PMIDs, and arXiv IDs to proper BibTeX format. It flags any reference it can't verify — which is your signal that the AI made it up.

When to use it: After any AI-generated writing. Before submitting anything. As the final quality gate in your pipeline.

The Pipeline: Chaining Them Together

Here's how these five tools work as a system:

Step 1: Landscape scan
  → openalex-database: "What's the field look like?"
  → Broad keyword search, citation trends, top authors

Step 2: Deep dive  
  → pubmed-database: "What's the clinical evidence?"
  → MeSH-controlled search, filter by study type

Step 3: Gap analysis
  → lit-synthesizer: "What's missing?"
  → Citation graph, coverage gaps, under-studied areas

Step 4: Synthesis
  → deep-research: "Write the review"
  → PRISMA-compliant systematic review with APA formatting

Step 5: Verification
  → citation-management: "Are these citations real?"
  → Cross-check every reference, generate clean BibTeX

Steps 1-3 take about 30 minutes. Step 4 takes 30-90 minutes. Step 5 takes 10 minutes. Total: under 3 hours for what used to take 3 weeks.

What's Still Missing

A few honest gaps:

Google Scholar access is limited. No official API exists, so skills that search Scholar rely on scraping — which is fragile and rate-limited. OpenAlex is the best workaround for most use cases.

Full-text analysis is rare. Most skills search titles, abstracts, and metadata. Few can read the actual paper content. The bgpt-paper-search skill (K-Dense) extracts 25+ structured fields from full texts, but requires a separate MCP server.

Cross-language search is weak. If important papers in your field are published in Chinese, Japanese, or Korean journals, most English-centric skills will miss them. AMiner's academic search is one of the few that handles Chinese-language scholarly content.

The "reading" part is still on you. These tools find, organize, and synthesize papers. But the intellectual work — understanding whether a study's methodology is sound, whether its conclusions follow from its data, whether it's relevant to your specific question — remains human work. AI compresses the logistics. It doesn't replace the thinking.

Last updated: April 3, 2026. All skills tested and actively maintained.