Best AI Agent Skills for Literature Review (2026)

If you've ever asked a chatbot for sources and gotten back three real papers, one paper that doesn't exist, and one real paper with the wrong authors, you already understand the problem these skills are trying to solve. The fix isn't a better model. It's giving the agent a search tool that hits an actual paper database, so citations come from a lookup instead of from memory.

That single design decision separates the skills worth installing from the ones that produce confident nonsense faster.

The shortlist

One note on the registry before you browse it yourself. The "Research" category is noisy. Several highly-starred entries filed under it are machine-learning research codebases rather than literature-review tooling, including Google DeepMind's research repo and Karpathy's autoresearch, which automates experiments on nanochat training runs. Both are excellent repositories and neither will help you write a related-work section. Star count is not a category signal.

Start with retrieval, not synthesis

claude mcp add paper-search-mcp -- npx -y openags/paper-search-mcp

2,302 stars, MIT, verified. It searches multiple academic sources and downloads the papers, which is the foundation everything else sits on. An agent that can fetch a PDF and quote from it is doing a different job from an agent recalling what it read during training.

arXiv MCP Server narrows to arXiv specifically, Apache-2.0, 3,009 stars, and it was updated on 29 July 2026, making it the most actively maintained option here. For physics, maths, CS, and quantitative biology, arXiv is where the preprints live and this is the cleaner tool. For medicine, social science, or anything where the canonical version sits behind a publisher, it will miss most of the field.

Install one of these before anything else. A synthesis skill with no retrieval underneath it is the hallucinated-bibliography machine you were trying to escape.

Paper Finder is narrower still, scoped to ML papers, 232 stars, and has no declared license. No license means no granted permission to use the work, which keeps it to personal machines regardless of how well it runs.

Synthesis, with the caveat that matters

GPT Researcher is the most-adopted option in this space at 28,753 stars, Apache-2.0, verified, updated mid-July 2026:

claude mcp add gpt-researcher -- npx -y assafelovic/gpt-researcher

It runs an autonomous loop: plan the question, search, read, and write a report with sources. It's strong at the first-pass survey, the one where you need to know what a field looks like before you can ask a sharper question. It is not built specifically for academic literature and will happily cite a blog post next to a journal article, so it's a scoping tool rather than a systematic-review tool.

Deep Research from 199 Biotechnologies takes the structured approach: an eight-phase pipeline that ends with citations, MIT licensed, 965 stars. The phased structure is the appeal, because it makes each step inspectable rather than handing you a finished report you have to reverse-engineer. It's community tier and unreviewed, so read it before running it. Its last update was April 2026.

AI Research Skills from Orchestra-Research is a library rather than one skill, MIT, verified, 11,292 stars. Worth browsing when your workflow spans research and implementation, since it covers engineering practice alongside investigation.

AcademicForge is the writing-side companion at 2,405 stars, aimed at academic prose and structure rather than discovery. Its license field reads "Other," which is the least useful possible answer. Check the repository directly before institutional use.

The part that stays manual

Retrieval fixes fabricated citations. It does not fix any of these:

Wrong-paper citations. The agent finds a real paper, summarizes it plausibly, and attributes a claim the paper never made. This is the dominant failure now that fabrication is mostly solved, and it's harder to catch, because everything in the bibliography checks out. Open the paper and find the sentence.

Search recall. These tools return what matched your query. A literature review is judged on what you missed, and no agent knows what it didn't retrieve. Vary your terms deliberately, including the terminology the field used a decade ago, and check the reference lists of the papers you did find.

Access. Most tools reach open-access and preprint sources. Paywalled literature is invisible to them, and in several fields that's the majority of what matters. A survey built only on what's freely downloadable has a systematic bias, and it will be the first thing a reviewer notices.

Recency versus consensus. Preprint servers surface the newest work, which is not the same as the accepted work. An arXiv-only survey overweights results nobody has replicated.

Duplicates that do not look like duplicates. The same work often exists as a preprint and a published version with a different title, a different author order, and a different DOI. Multi-source search returns both, and an agent counting papers will count them twice. If a tool reports that it found 214 studies, that number is a search result, not a corpus, and the gap between the two is usually 20 to 40 percent.

Our PDF processing skills guide covers the mechanics of getting text out of the papers once you have them, which matters more than people expect when the PDF is a scan of a 1997 conference proceeding.

On autonomous paper writing

Several registry entries promise the full pipeline from research idea to finished paper. AutoResearchClaw is the most-starred at 13,930, MIT, verified tier, and actively maintained.

I'd treat the category as a research demo rather than a workflow, for a reason that has nothing to do with code quality. Journals and conferences increasingly require disclosure of AI use, several prohibit AI-generated text outright, and authorship carries accountability for every claim in the paper. A tool that produces a submittable draft produces something you're personally answerable for and didn't verify.

The interesting use is different: run it, then read what it got wrong. The gap between its draft and a defensible one is a fast map of where the hard parts of your problem sit.

What to run

Scoping a new field: GPT Researcher for the survey, then Paper Search MCP to pull the papers it cited so you can check them.

Systematic review: Paper Search MCP or arXiv MCP for retrieval, PDF Tools for extraction, and your own protocol for inclusion criteria. Do not let a skill decide what's in scope. That decision is the review.

Writing up: AcademicForge for structure, plus DOCX if your submission target wants Word rather than LaTeX.

ML and CS specifically: arXiv MCP alone covers most of the ground.

Two installs, one for retrieval and one for synthesis, is the whole stack. Researchers looking at the GUI side of this should see our academic research tools roundup and the literature review tools comparison, and the researchers list covers the broader workflow. Before installing anything community-tier, our piece on author trust tiers explains what the labels do and don't promise.

FAQ

Will these stop Claude from making up citations? Mostly, and only if you install a retrieval tool. A search-backed skill returns papers that exist because it looked them up. What survives is misattribution, where a real paper gets credited with a claim it never made, so spot-check the actual sentences behind your key citations.

Can these access paywalled journals? No. They reach open-access repositories, preprint servers, and metadata. If your institution has subscriptions, those live behind an authenticated proxy these tools do not use. Treat any agent-built bibliography as a starting set that is systematically biased toward what is free.

Is it safe to use AI for a literature review I'll publish? For discovery and organization, that's ordinary tool use. For generated prose, check your target venue's policy first, because requirements vary from full disclosure to outright prohibition, and they've been changing. You remain accountable for every claim regardless of what produced the draft.

Which single one should I install? Paper Search MCP, for breadth of sources, MIT licensing, and verified tier. If you work only in physics, maths, or CS, arXiv MCP Server is the better fit and is more actively maintained.

Why do some highly-starred "research" skills seem unrelated to literature review? Because registry categories are assigned loosely and machine-learning research code lands in the same bucket as research assistants. A repo with 90,000 stars for automating training experiments will outrank a purpose-built paper-search tool on every popularity sort. Read the description, not the ranking.