Skip to content

Phở Founder Lifetime: $59 once, 1 year of Phở AI included. Limited seats. See the offer

· The Phở team

Best MCP servers for research in Claude and ChatGPT (2026)

Seven research connectors for Claude, ChatGPT and other AI assistants, compared on what each one can actually do: Scite, Elicit, PubMed, alphaXiv, bioRxiv, ClinicalTrials.gov and Phở Research Tools. Tools checked 4 November 2026.

On this page

Tools checked: 2026-11-04. We connected each server to an MCP client and read the tools it exposes. Prices and plan limits change, so for anything we could not read on the vendor’s own page we point you to that page instead of quoting a number.

A research MCP server turns an AI assistant from something that remembers papers into something that looks them up. That matters because assistants still invent references, and a lookup is the cheapest fix. This page compares seven connectors that researchers actually install, including ours, Phở Research Tools. We built one of the seven, so read the table as a list of what each tool does, and judge us by the same standard.

At a glance

ConnectorBest atMain tools it exposesAccount
SciteHow a paper has been cited since, and checking papers against citation contextsearch_literature, verify_papers, citation_graph, read_fulltext, bibliographyScite account; see scite.ai for plans
ElicitRunning Elicit’s systematic review and report workflow from chatsearch_papers, search_trials, create_systematic_review, create_report, library toolsElicit account; see elicit.com/pricing
PubMedBiomedical search straight from NCBI, with PubMed Central full textsearch_articles, get_full_text_article, find_related_articles, get_article_metadataNone for the data; connector from Claude’s directory
alphaXivarXiv papers in CS, maths, physics and statisticsdiscover_papers, get_paper_content, answer_pdf_queries, library toolsalphaXiv account for library features
bioRxivRecent biology and medicine preprints by date and categorysearch_preprints, get_preprint, search_published_preprints, search_by_funderNone for the data
ClinicalTrials.govTrial registry search by condition, sponsor, eligibility and endpointssearch_trials, get_trial_details, search_by_sponsor, analyze_endpoints, search_by_eligibilityNone for the data
Phở Research ToolsChecking reference lists and retractions, evidence level, PRISMA and screening logsverify_references, check_retractions, search_papers, find_evidence, format_citations, prisma_flow_diagram, screening_log, find_datasets, find_tools, load_datasetNone to start: 30 reference checks a day

Scite

Scite’s strength is citation context. Its Smart Citations show the sentence in which a later paper cites an earlier one and whether that statement supports or contrasts it. Through MCP, search_literature returns those snippets, citation_graph walks from a seed paper to its references or to the work that cited it, verify_papers checks DOIs you pass it for retractions and contrasting evidence, and read_fulltext reads open-access and licensed text. If your question is “has this finding held up?”, Scite is the strongest tool here.

Pick Scite if you need to know how a paper has been received, or want full-text excerpts as evidence. It needs a Scite account; check scite.ai for what each plan includes.

Elicit

Elicit’s connector brings its own workflow into the assistant: search_papers and search_trials over its corpus of 138 million papers, create_systematic_review to start a screening and extraction run, create_report for a research report, and library tools to save and organise sources. The work runs inside Elicit, so you get Elicit’s limits (5,000 papers screened on Pro, up to 40,000 on Enterprise) and its audit trail.

Pick Elicit if your team already runs reviews in Elicit and you want to start or check them from chat. Pro is $49 per user per month billed annually (pricing page, 2026-10-29).

PubMed

The PubMed connector queries NCBI directly. search_articles takes PubMed’s own query syntax, get_full_text_article pulls the text of open papers from PubMed Central, find_related_articles follows PubMed’s similar-articles links, and get_article_metadata returns authors, journal, MeSH terms and identifiers. It covers biomedicine only, but within that it is the most direct source there is.

Pick PubMed if you work in medicine or life sciences and want search with no account and no middle layer.

alphaXiv

alphaXiv covers arXiv: computer science, maths, physics, statistics and quantitative biology, but not PubMed-style biomedical journals. discover_papers searches, get_paper_content and answer_pdf_queries read a paper and answer questions about it, and there are tools for your alphaXiv library and for finding researchers.

Pick alphaXiv if your field publishes on arXiv first.

bioRxiv

The bioRxiv connector covers bioRxiv and medRxiv preprints. Worth knowing before you install it: search_preprints filters by date range and subject category, with no keyword or author search, so it is best for “what was posted in neuroscience this month”, not for finding a specific paper. search_published_preprints tells you whether a preprint was later published in a journal. Everything it returns is a preprint that has not been peer reviewed.

Pick bioRxiv if you track new preprints in a field, or need to check a preprint’s publication status.

ClinicalTrials.gov

This connector wraps the ClinicalTrials.gov API. Beyond search_trials and get_trial_details it can list a sponsor’s pipeline, find investigators and sites, compare endpoints across similar trials, and match trials to eligibility criteria.

Pick ClinicalTrials.gov if your question is about trials that are planned, running or finished, rather than published papers.

Phở Research Tools

This one is ours. Phở Research Tools is built around a question the others answer only in part: is this reference real, and is it still standing? verify_references checks a whole reference list against Crossref, PubMed and OpenAlex and marks each entry verified, not found or likely fabricated (a DOI that is not registered or points to another paper, or a paper missing from a journal these databases index). A paper in a journal they do not cover comes back as not found, never as fabricated. check_retractions looks up retraction and correction notices. find_evidence lists the strongest evidence on a claim with its level of evidence and each paper’s conclusion quoted verbatim, and does not tell you whether the paper supports you. format_citations, prisma_flow_diagram and screening_log produce a bibliography, a PRISMA 2020 diagram and a screening log with counts. find_datasets, find_tools and load_dataset find and open open datasets and research software with their licences.

What it does not do: it does not read full text, it does not classify citation context the way Scite does, and it does not run a systematic review for you.

Pick Phở Research Tools if you want to check a manuscript’s references or a student’s thesis before it goes out, at no cost. It works without an account (30 reference checks and 200 tool calls a day); a free Phở account gives 1,000 reference checks a month, and Pro or Founder Lifetime 3,000. The endpoint is https://mcp.pho.chat/mcp, the setup guide for Claude, ChatGPT, Cursor, Claude Code, Gemini CLI and others is at pho.chat/tools/mcp, and the code is open source under the MIT licence at github.com/thaohienhomes/pho-research-tools.

Not in the table

Consensus also lists a connector for AI assistants. We could not connect to it on 4 November 2026, so we have not read its tools and leave it out rather than guess. We will add it once we have.

How to choose

  • Checking references before you submit: Phở Research Tools first, then Scite verify_papers for citation context on the ones that matter.
  • Biomedical literature search: PubMed, plus bioRxiv for preprints and ClinicalTrials.gov for trials.
  • Computer science and physics: alphaXiv.
  • Systematic reviews: Elicit if you review in Elicit; Phở screening_log and prisma_flow_diagram for the PRISMA paperwork wherever you screened.
  • Has this finding held up? Scite.

An honest note: we build Phở Research Tools, one of the seven. The tool lists above are what each server exposed to our MCP client on 4 November 2026; servers add and rename tools often.

Sources

  1. Tool lists read from each MCP server through an MCP client on 4 November 2026: Scite, Elicit, PubMed (Claude directory), alphaXiv, bioRxiv, ClinicalTrials.gov, Phở Research Tools.
  2. Elicit, Pricing. https://elicit.com/pricing (checked 2026-10-29)
  3. Phở Research Tools, setup guide and limits. https://pho.chat/tools/mcp (checked 2026-11-04)
  4. Phở Research Tools, source code. https://github.com/thaohienhomes/pho-research-tools

Spot an error on this comparison, including about your own product? Email support@pho.chat. We fix it and note the date.

Frequently asked questions

What is an MCP server for research?

MCP (Model Context Protocol) is the standard way to plug outside tools into an AI assistant such as Claude, ChatGPT, Cursor or Gemini CLI. A research MCP server gives the assistant real lookups, for example searching PubMed or checking a DOI, so it answers from a database instead of from memory.

Which research MCP server should I install first?

It depends on the job. For checking whether a reference list is real, Phở Research Tools or Scite. For how a paper has been cited since, Scite. For biomedical search with full text from PubMed Central, the PubMed connector. For running a systematic review inside Elicit from chat, Elicit. For trial registries, ClinicalTrials.gov. Most people end up with two or three.

Is there a free MCP server that checks if references are real?

Yes. Phở Research Tools works without an account: 30 reference checks and 200 tool calls a day. A free Phở account raises that to 1,000 reference checks a month. It checks each reference against Crossref, PubMed and OpenAlex and flags DOIs that are not registered or point to another paper. The PubMed, bioRxiv and ClinicalTrials.gov connectors only read free public databases too, but they search rather than check a list.

Can these connectors tell me whether a paper supports my claim?

Partly. Scite classifies citation statements as supporting or contrasting and can verify papers you pass it. Phở find_evidence quotes each paper's own conclusion with its level of evidence, and leaves the judgment to you. No connector here reads every paper in full and decides for you, and you should be wary of one that claims to.

Read next