Best AI Agent Skills for Researchers in 2026

A PhD candidate at a UK research university told me she stopped using Zotero for two months in early 2026. Not because Zotero broke, because she replaced the workflow with three Claude Code skills running in series. Search the literature, retrieve full PDFs, summarize and extract citations into a BibTeX file. The skills cost nothing (open source). Her supervisor noticed the output, not the tooling change.

This is the actual state of AI for researchers in 2026: not a single product replacing Zotero, but a toolkit of installable skills that handle slices of the research workflow. Below are the ten I'd install for any researcher starting fresh today, with install commands and what each one is good for.

Already on the site: see our full list for researchers for tool reviews and the broader catalog.

Quick comparison

Literature search and retrieval

1. deep-research, start here

The deep-research skill is the closest thing to a built-in literature review assistant. Give it a research question, it generates 8-15 sub-queries, runs them in parallel against Google, Semantic Scholar, and arXiv, deduplicates results, and produces a structured summary with inline citations. For a graduate student writing a thesis chapter, the time saved on the initial scoping phase is substantial, a 6-hour Sunday afternoon search becomes a 45-minute review of structured output.

Limitations: it's still searching open web and open-access sources. Paywalled journal content (Nature, Cell, ScienceDirect) won't be in the corpus unless your institution has unlocked access proxies, and the skill doesn't currently route through institutional logins.

git clone https://github.com/<author>/deep-research ~/.claude/skills/deep-research

For paywalled access, pair this with your university's VPN and the playwright-mcp skill for authenticated page rendering.

2. arxiv-mcp-server, arXiv specialist

If your field lives on arXiv (CS, physics, math, statistics), the arxiv-mcp-server is essential. It exposes arXiv's API as MCP tools that Claude can call: search by topic, fetch abstracts, retrieve full PDFs, get author publication histories. Unlike the deep-research skill (which is generalist), this one returns the arXiv-specific metadata you need for proper citation: identifier, version, submission date, primary category.

{
  "mcpServers": {
    "arxiv": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/arxiv-mcp-server"]
    }
  }
}

Add to your ~/.claude/settings.json and restart Claude Code.

3. paper-search-mcp, multi-database search

For interdisciplinary work that needs to cross arXiv, PubMed, Semantic Scholar, and bioRxiv, paper-search-mcp aggregates the queries. One prompt, results from all five databases, deduplicated by DOI. Output includes BibTeX-ready citation strings.

npm install -g paper-search-mcp

Then add to your MCP config. The full repo is at github.com/<author>/paper-search-mcp.

Document handling

4. pdf-tools, beyond text extraction

Most PDF parsing tools handle text but break on tables, figures, and scanned pages. The pdf-tools skill uses pdfplumber + tabula-py + Tesseract OCR in series, which means you can extract a table from a 1992 scanned PhD thesis and get a clean DataFrame. This matters for systematic reviews and meta-analyses where the source data is locked inside published PDFs.

git clone https://github.com/<author>/pdf-tools ~/.claude/skills/pdf-tools
pip install -r ~/.claude/skills/pdf-tools/requirements.txt

5. xlsx-spreadsheet, data analysis

If your research involves any quantitative analysis (extracting numbers from a clinical trials registry, building a meta-analysis dataset, analyzing survey responses), xlsx-spreadsheet wraps openpyxl + pandas to let Claude read, filter, pivot, and chart your data without you switching to a notebook. For researchers who think in tables more than code, this is the productivity unlock.

Pairs with pdf-tools: extract tables from PDFs with pdf-tools, hand them to xlsx-spreadsheet for cleaning and analysis.

6. docx-word, write the paper from research

Once your research is done, docx-word generates Word documents with tables, figures, citation styles (APA, Chicago, MLA), and bibliography sections. It supports tracked changes for collaborative review. Researchers in fields where Word is the publishing standard (medicine, social sciences, humanities) use this for first-draft generation.

Web research

7. firecrawl-skill, grey literature

Grey literature (think tank reports, government white papers, NGO publications) often lives outside academic databases. Firecrawl is the most reliable scraper I've tested for these use cases, it handles JavaScript-rendered pages, paginates correctly through long reports, and respects robots.txt. For researchers studying policy, public health, or emerging technology, the grey literature in your literature review just got 10x more accessible.

npm install -g firecrawl-mcp

Requires a Firecrawl API key (free tier covers ~500 pages/month).

8. playwright-mcp, interactive page rendering

Some research data lives behind authentication or interactive UIs (think a clinical trials database with a multi-step search form, or an institutional repository with a custom search). playwright-mcp lets Claude drive a real browser, fill in forms, click through pagination, and screenshot results. For one-off retrieval tasks where writing a custom scraper would take longer than the data is worth, this is the right tool.

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["@playwright/mcp@latest"]
    }
  }
}

9. claude-seo, keyword discovery for research questions

This sounds like it's for marketers, but I've seen researchers use claude-seo's keyword research mode for a specific purpose: finding the actual terminology a community uses for a phenomenon. If you're studying "informal caregiving in dementia families," what do practitioners search for online? claude-seo's keyword cluster output will surface "respite care," "dementia caregiver burnout," and "carer support" — the language your participants and target communities use, which you can then bring back into your interview guide or survey design.

See our claude-seo review for the full feature breakdown.

Building your own

10. skill-creator, build research-specific skills

Every research workflow has 1-2 idiosyncratic steps that no public skill solves. A computational biologist needs a skill that understands FASTA files. A social scientist running interviews needs a skill that converts audio to coded transcripts in NVivo format. The skill-creator skill helps you write your own, it scaffolds a SKILL.md file, structures the references folder, and validates the metadata.

git clone https://github.com/anthropics/skills/tree/main/skill-creator

For a tutorial walkthrough, see How to install skill-creator.

A realistic research stack

For a PhD student in social sciences starting a thesis:

1. deep-research, initial literature scoping (free) 2. paper-search-mcp, systematic search across databases (free) 3. pdf-tools, full-text extraction from retrieved PDFs (free) 4. firecrawl-skill, grey literature scraping (free for low volume, $20/mo for heavy use) 5. xlsx-spreadsheet, data extraction and analysis (free) 6. docx-word, first draft and citation management (free)

Total monthly cost: $0-20 depending on Firecrawl usage. Compare to a Zotero + Atlas.ti + EndNote stack at $150-400/year, plus the labor savings on literature scoping and PDF parsing.

Security considerations

Three things to check before installing any community skill:

Read the SKILL.md and any code files. Most skills are short, under 500 lines total. Skim them.

Check the GitHub repo's recent commit history. A skill with 6K stars that hasn't been updated in 12 months is a maintenance risk. Active development = maintained.

For MCP servers, audit what they connect to. The arxiv-mcp-server only talks to arXiv's API. The firecrawl-skill talks to Firecrawl's API and any URL you give it. Know what's leaving your machine.

Honest limitations

The skills above are good. They're not Zotero. Specifically:

Reference management, keeping a long-term, manually-curated bibliography across years of work, is still better in Zotero or Mendeley. The skills handle one-shot retrieval beautifully but don't replace a citation manager.

PDF annotation (highlighting passages, taking margin notes) is not a skill workflow. If your research method involves close reading and annotation, keep using Zotero or PDF Expert for that step.

Real-time collaboration with co-authors. Skills are local, single-user. For collaborative writing, Overleaf or Google Docs still win.

The right framing: skills handle the search, retrieval, and parsing layer. Your existing reference manager handles the long-term knowledge layer. Together, your 6-hour Sunday afternoon search becomes 45 minutes.

FAQ

Are these skills safe to install?

The 10 skills above are open source on GitHub with thousands of stars and active maintenance. Read the code (most are under 500 lines), check the issue tracker, and verify the GitHub link before installing. See our skill audit guide for a deeper checklist.

Do these work outside Claude Code?

Most are Claude Code-specific (skills + MCP servers). The MCP servers (arxiv-mcp-server, firecrawl-skill, playwright-mcp) work with any MCP-compatible client, including Claude Desktop and some Anthropic-compatible IDE plugins.

Can I use these for confidential research data?

Skills run locally; the file processing happens on your machine. The web search and paper retrieval skills (deep-research, firecrawl, paper-search-mcp) send queries to external APIs, so don't use them on confidential queries. For local-only workflows, pdf-tools, xlsx-spreadsheet, and docx-word are safe.

What's the difference between MCP servers and Claude Skills?

MCP servers expose APIs; skills are workflow templates. See our explainer on MCP servers vs Agent Skills for the practical difference.

Will these replace my research assistant?

For literature scoping and data parsing, mostly yes. For the human work, running interviews, designing studies, interpreting results, writing the paper's argument, no. The skills speed up the boring 60% of research work so you can spend more time on the 40% that matters.