AI research is bottlenecked by literature review, dataset discovery, and experiment tracking. MCPs cut all three: arXiv + Perplexity for paper discovery, HuggingFace for datasets and models, Pinecone or Supabase for embedding your own paper library, Jupyter for experiment execution, and GitHub for code. One Claude session replaces the usual 6-tool workflow.
A research scientist, PhD student, or applied ML engineer reading 5–20 papers a week, running experiments, and publishing insights.
Literature review on steroids — 'summarise the last 6 months of work on sparse attention with citations'. Cuts lit review from days to hours.
Breadth search across arXiv, blog posts, and release notes. Useful for 'is there already a paper doing X?' before you invest in a research direction.
Scrape full paper PDFs (arXiv, openreview, papers with code) into markdown. Extract tables, equations, and related work from any paper.
Your personal paper library with pgvector. Claude embeds every paper you've read and answers 'have we seen this technique before?' against your own history.
Clone repos referenced in papers, read the implementation, open PRs on your own research code. Critical for reproducibility.
Your research journal. Claude writes paper summaries, experiment logs, and idea threads directly into the right Notion databases.
Experiment tracking — hyperparameters, eval metrics, checkpoints — all in Postgres. Claude queries your runs and compares trajectories without opening MLflow.
You're writing a related work section on long-context transformers. Claude uses Tavily to find the 30 most-cited recent papers on the topic, Perplexity to summarise the top 10 with key contributions, Firecrawl to pull full PDFs, Supabase pgvector to check which of these you've already read (it knows from your personal embedding library), and Notion to write the draft related work section. You get a structured 2-page draft with 30 proper citations in under an hour, where it used to take 2 days.
Researchers save 12–18h/week on literature review, paper summarisation, and related work. Downstream: you read 2x more papers and ship papers 2x faster.
Paste a research topic in Notion and an agent uses Perplexity to gather sources, summarize findings, and structure them.
Schedule a Firecrawl scrape of any website and store the structured results directly in a Supabase table for analysis.
Run Tavily searches on scheduled topics and index the results in Supabase for trend analysis and content research.
Run daily Perplexity searches on competitors and log product updates, pricing changes, and news to a Notion tracker.
Parse your GitHub repos and build a Neo4j knowledge graph of files, functions, imports, and authors for code intelligence.
A community-maintained arXiv MCP exists with search, download, and metadata tools. It's thin — most heavy lifting is done by Perplexity (discovery) + Firecrawl (PDF extraction). arXiv's API is open, so any MCP works.
The HuggingFace MCP covers models, datasets, and spaces. Critical for ML engineers: 'find me a text classification dataset with at least 100K examples in Portuguese' → answer in seconds.
Yes — the Jupyter MCP launches and executes notebook cells. Combined with the GitHub MCP (clone repo) and HuggingFace MCP (download model), you can reproduce most papers end-to-end from a single Claude session.
For <100K papers, no. Supabase+pgvector is free up to 500MB and fast enough. Pinecone's advantage kicks in at 1M+ embeddings where latency and sharding matter. Most individual researchers never hit that.
It can draft reviews — but the final review must be yours. Most reviewers use Claude for: initial summarisation, identifying key contributions, checking citations for completeness, and drafting the 'strengths/weaknesses' sections. The judgment stays human.
A technical founder (0–10 employees) building a B2B SaaS who ships code, handles billing, writes marketing, and answers support — all in the same day.
An indie hacker with a Twitter audience, a newsletter, 1–3 shipped products, and zero employees. Ships daily, markets constantly, avoids meetings.
A developer building AI agents, chatbots, or autonomous workflows. Needs search, scraping, vector storage, and LLM orchestration — all as tools the agent can call.
Install the full stack in one command, or cherry-pick the MCPs you need.
Browse all MCPs