Guides
MCP servers for research data: arXiv, PubMed, Semantic Scholar, OpenAlex, SEC EDGAR, FRED, ClinicalTrials.gov, Crunchbase and PitchBook
Most of these sources now have an official MCP server, and the ones that do not have a maintained open-source one. Anthropic runs free, no-sign-in connectors for PubMed and ClinicalTrials.gov; OpenAlex, FRED, Crunchbase and PitchBook publish their own remote servers that you sign in to; Ai2 runs one for the Semantic Scholar corpus. arXiv and SEC EDGAR have no official server, and the open-source servers blazickjp/arxiv-mcp-server and dgunning/edgartools are the most used. Google Scholar has no API at all, so every Google Scholar MCP server reads its web pages, against its robots.txt; use OpenAlex or Semantic Scholar instead.
Everything below was checked on November 15, 2026 against the official MCP registry, the Claude connector directory, each source’s own documentation and each GitHub repository. Star counts are from that day.
Research MCP servers compared
| Source | Official or first-party server | Best open-source option (stars) | Sign-in or key | Upstream limit | Notes |
|---|---|---|---|---|---|
| arXiv | None | blazickjp/arxiv-mcp-server (3,281) | None | 1 request every 3 seconds | Reads the LaTeX source by section |
| PubMed | Anthropic’s PubMed connector for Claude | cyanheads/pubmed-mcp-server (168) | None; optional NCBI key | 3 requests/s, 10 with a key | Full text only for PubMed Central articles |
| ClinicalTrials.gov | Anthropic’s Clinical Trials connector for Claude | cyanheads/clinicaltrialsgov-mcp-server (101) | None | None published | Uses the v2 API |
| Semantic Scholar | Ai2’s Asta Scientific Corpus Tool | zongmin-yu/semantic-scholar-fastmcp-mcp-server (166) | Optional key | 1,000 requests/s shared by all keyless users; 1/s on a new key | Key raises the Asta limit |
| OpenAlex | OpenAlex’s own server, mcp.openalex.org |
cyanheads/openalex-mcp-server (18) | Free OpenAlex account (official server) | 10 requests/s keyless, 30 with a free key | Free key: $1 of API use a day |
| Google Scholar | None, and no API | JackKuo666/Google-Scholar-MCP-Server (433) | None | robots.txt disallows /scholar |
Last commit March 2025; use OpenAlex instead |
| SEC EDGAR | None from the SEC | dgunning/edgartools (2,855) | Your name and email in the User-Agent | 10 requests/s | stefanoamorelli/sec-edgar-mcp (387) is AGPL |
| FRED | St. Louis Fed’s FRED MCP Connector | stefanoamorelli/fred-mcp-server (134) | Connector account (official); FRED API key (open source) | 120 requests/min per key | Official server needs no FRED API key |
| Crunchbase | Crunchbase MCP, mcp.crunchbase.com |
None worth using | Crunchbase account | Data tokens on your plan | Listed in Claude and ChatGPT directories |
| PitchBook | PitchBook Premium connector | None | PitchBook license and SSO | Not published | Seat-based, unlimited or trial license |
“Official” here means run by the data source itself or by the AI vendor whose client it is built into. The MCP registry is in preview and checks only that a publisher owns the namespace it publishes under, such as org.openalex or io.github.<user>; it does not review the code, so treat an open-source server as code you are choosing to run.
How to add a research MCP server to Claude, Cursor and ChatGPT
The client decides how you add a server, and the server decides whether you need anything installed.
- A remote server (a URL) needs nothing installed. In claude.ai and Claude Desktop, add it as a custom connector under Settings > Connectors. Anthropic says Free plans are limited to one custom connector. In Claude Code, run
claude mcp add --transport http <name> <url>. - A local server (a
uvx,npxordockercommand) runs on your machine. Claude Desktop reads it fromclaude_desktop_config.json(~/Library/Application Support/Claude/on macOS,%APPDATA%\Claude\on Windows). Claude Code adds it withclaude mcp add <name> -- <command>. - Cursor reads both kinds from
~/.cursor/mcp.json(every project) or.cursor/mcp.json(one project). A local server usescommandandargs; a remote one usesurl. - ChatGPT only connects to a server reachable at a public HTTPS address (or through OpenAI’s Secure MCP Tunnel), so a local
uvxserver will not work there as is. Use a directory plugin where one exists, or add a remote server from the Plugins page in developer mode if your plan and workspace admin allow it.
A local server takes the same JSON entry in claude_desktop_config.json and Cursor’s mcp.json:
{ "mcpServers": { "arxiv": { "command": "uvx", "args": ["arxiv-mcp-server"] } }}In Cursor, a remote server is one line in the same block, such as "openalex": { "url": "https://mcp.openalex.org/mcp" }. In Claude Desktop, add remote servers through Settings > Connectors.
arXiv MCP server
arXiv has no official MCP server, and the one to use is blazickjp/arxiv-mcp-server: 3,281 stars, Apache-2.0, last commit November 14, 2026, listed in the official MCP registry at version 0.8.2. It exposes 19 tools. Beyond search and download, it reads a paper’s original LaTeX one section at a time, exports BibTeX from arXiv’s own metadata, keeps topic watches, and builds a citation graph through the Semantic Scholar API. Papers are saved to ~/.arxiv-mcp-server/papers.
claude mcp add --transport stdio --scope user arxiv -- uvx arxiv-mcp-serverFor Claude Desktop or Cursor, use the arxiv entry in the JSON above. Install it with uvx from PyPI: the project warns that an unrelated npm package uses the same name.
Other options:
- cyanheads/arxiv-mcp-server (7 stars, TypeScript, 4 tools) is smaller, and its author runs a public hosted copy at
https://arxiv.caseyjhand.com/mcp, which is the easy route into ChatGPT. - Claude Science includes arXiv in its built-in Literature Graph connector, alongside OpenAlex.
- The page that ranked first on Google for “arxiv mcp” on November 15, 2026 is a directory listing of the blazickjp project built from its early README. It calls MCP the “Message Control Protocol” and lists five features from before the project grew to 19 tools. Read the repository, not the listing.
arXiv’s API terms ask for no more than one request every three seconds, one connection at a time, across all the machines you control. The descriptive metadata is CC0; each paper keeps its own licence, and the terms say not to store and serve the PDFs from your own servers without the copyright holder’s permission.
PubMed MCP server and the Claude PubMed connector
Claude has an official PubMed connector, made by Anthropic and added to the directory in November 2025. It needs no sign-in, and the Claude Academy tutorial says it is free for all Claude users. Its seven tools search PubMed, read metadata and abstracts, fetch full text when the article is in PubMed Central, find related articles, match a citation to a PMID, convert between PMID, PMCID and DOI, and check copyright status.
Add it from Settings > Connectors in Claude. In Claude Code, the tutorial’s route is to connect it in Claude and sign in to Claude Code with the same account; it then shows under /mcp. The endpoint itself is public, and it answered an MCP initialize and tool list with no key on November 15, 2026:
claude mcp add --transport http pubmed https://pubmed.mcp.claude.com/mcpOutside Claude, use cyanheads/pubmed-mcp-server: 168 stars, Apache-2.0, last commit November 13, 2026. It has 12 tools and covers more than the official connector: Europe PMC (preprints and patents included), full text that falls back from PubMed Central to Europe PMC to Unpaywall, NIH iCite citation metrics, MeSH lookup, and citations formatted in APA, MLA, BibTeX, RIS or Vancouver with retraction notices attached.
claude mcp add pubmed-mcp-server -e NCBI_API_KEY=your-key -- npx -y @cyanheads/pubmed-mcp-serverDrop -e NCBI_API_KEY=... to run without a key. The author also runs a hosted copy at https://pubmed.caseyjhand.com/mcp. Two older servers still rank on Google: JackKuo666/PubMed-MCP-Server (132 stars, last commit May 2025) and grll/pubmedmcp (123 stars, last commit September 2025).
Every PubMed server calls NCBI’s E-utilities, which allow three requests a second from one IP address without a key and 10 a second with a free NCBI API key. For the papers behind a PMID or DOI, see legal open-access full text for agents, and before citing anything, check whether it was retracted.
ClinicalTrials.gov MCP server
Claude also has an official Clinical Trials connector, made by Anthropic and added in January 2026, with no sign-in. Its six tools search trials, read one trial’s details, search by sponsor, by investigator and by eligibility, and analyze endpoints. It is at https://hcls.mcp.claude.com/clinical_trials/mcp, which also answered keyless on November 15, 2026:
claude mcp add --transport http clinical-trials https://hcls.mcp.claude.com/clinical_trials/mcpThe open-source option is cyanheads/clinicaltrialsgov-mcp-server: 101 stars, Apache-2.0, last commit November 13, 2026, 7 tools, no key. It fetches posted results (outcomes, adverse events, participant flow) and counts matching studies without downloading them. Local install:
{ "mcpServers": { "clinicaltrialsgov-mcp-server": { "command": "npx", "args": ["-y", "clinicaltrialsgov-mcp-server@latest"] } }}A hosted copy runs at https://clinicaltrials.caseyjhand.com/mcp. ClinicalTrials.gov publishes no rate limit for its API. What the API accepts, and the classic API’s retirement in 2024, are in the ClinicalTrials.gov API v2 guide.
Semantic Scholar MCP server
Ai2, which runs Semantic Scholar, publishes the Asta Scientific Corpus Tool, a remote MCP server over the Semantic Scholar paper graph at https://asta-tools.allen.ai/mcp/v1. Ai2 describes the graph as more than 225 million papers and 2.4 billion citation edges, plus a snippet search over 12 million full-text papers. Its eight tools search papers by relevance or title, fetch one paper or a batch, list citations, search authors and their papers, and search snippets. It answered keyless on November 15, 2026; a free key, requested through Ai2’s form and sent as an x-api-key header, raises the rate limit.
claude mcp add --transport http asta https://asta-tools.allen.ai/mcp/v1 --header "x-api-key: YOUR_KEY"Two open-source servers wrap the public Semantic Scholar API:
- zongmin-yu/semantic-scholar-fastmcp-mcp-server: 166 stars, MIT, last commit March 20, 2026. Run it with
uvx semantic-scholar-fastmcpand setSEMANTIC_SCHOLAR_API_KEYin itsenvblock. - smaniches/semantic-scholar-mcp: 26 stars, MIT, last commit November 12, 2026, 14 tools, no key needed to start:
claude mcp add semantic-scholar -- uvx s2-mcp-server.
Without a key, the Semantic Scholar API allows 1,000 requests a second shared among every unauthenticated user, so a busy moment can throttle you; a new key starts at one request a second on all endpoints. The API licence lets Ai2 end access at any time.
OpenAlex MCP server
OpenAlex runs its own server at https://mcp.openalex.org/mcp, listed in the official MCP registry as org.openalex/openalex. You sign in with a free OpenAlex account, and every query runs on your own API key and daily budget: $1 of API use a day on a free key. Its 10 tools are read-only. They search works by keyword or by meaning, fetch a work by OpenAlex ID, DOI, PMID or PMCID, check up to 25 references at once for whether they exist, list citations, profile authors, institutions, journals and funders, and count or analyze any set of works. Each works query returns the exact OpenAlex query it ran, so a search can be rerun or cited in a methods section.
claude mcp add --transport http openalex https://mcp.openalex.org/mcpOpenAlex says the connector is in Claude’s connector directory on every plan, including Free, and that a ChatGPT version is coming. The open-source alternative is cyanheads/openalex-mcp-server (18 stars, Apache-2.0), which works keyless with an optional OPENALEX_API_KEY.
The OpenAlex API itself allows 10 requests a second without a key, 30 with a free key and 50 on an annual plan; semantic search is limited to one a second. A free key raises the daily budget tenfold and is required for cursor paging past the first page. The data is CC0.
Google Scholar MCP server
Google Scholar has no API and no official MCP server, so a Google Scholar MCP server is a scraper. Google’s help page answers requests for bulk access with “Sorry, we’re unable to provide bulk access,” and Scholar’s robots.txt disallows /scholar, the search path. The most-starred server, JackKuo666/Google-Scholar-MCP-Server (433 stars, no licence file, last commit March 25, 2025), uses the scholarly Python package, which reads Scholar’s pages. Expect it to be blocked under load.
For what people usually want from Scholar (find papers on a topic, count citations, follow who cites whom), the OpenAlex and Semantic Scholar servers above do the same job through APIs that allow it.
SEC EDGAR MCP server
The SEC publishes no MCP server; every SEC EDGAR server in the registry is third-party. Two open-source projects lead:
- dgunning/edgartools: 2,855 stars, MIT, last commit November 15, 2026. A Python library for filings and XBRL financials that ships a local MCP server in its
aiextra. - stefanoamorelli/sec-edgar-mcp: 387 stars, AGPL-3.0 (with a commercial licence on request), last commit October 3, 2026. Tools for company lookup, 10-K, 10-Q and 8-K sections, XBRL statements and Form 3, 4 and 5 insider trades, and every response includes the filing URL.
edgartools in Claude Desktop or Cursor:
{ "mcpServers": { "edgartools": { "command": "uvx", "args": ["--from", "edgartools[ai]", "edgartools-mcp"], "env": { "EDGAR_IDENTITY": "Your Name your.email@example.com" } } }}sec-edgar-mcp runs from Docker:
{ "mcpServers": { "sec-edgar-mcp": { "command": "docker", "args": ["run", "-i", "--rm", "-e", "SEC_EDGAR_USER_AGENT=Your Name (your@email.com)", "stefanoamorelli/sec-edgar-mcp:latest"] } }}cyanheads/secedgar-mcp-server (12 stars) adds 13F and 13D/13G ownership and has a hosted copy at https://secedgar.caseyjhand.com/mcp.
Whichever you pick, put your own name and email in its User-Agent setting. The SEC asks every automated client to declare one and caps each user at 10 requests a second across all machines; a request without one gets a 403. The SEC EDGAR API in Python guide covers the endpoints these servers call.
FRED MCP server
The Federal Reserve Bank of St. Louis runs an official FRED MCP Connector at https://mcp.stlouisfed.org. Its help page says no FRED API key is required; you sign in with a separate FRED MCP Connector account, not your FRED account, and approve the assistant. Its five tools search series, read metadata and observations, make a CSV or JSON download link, and draw a graph of up to five series. The page lists FRED in Claude’s connector directory and says any assistant that takes a custom MCP connector can use the URL:
claude mcp add --transport http fred https://mcp.stlouisfed.orgThe open-source option is stefanoamorelli/fred-mcp-server: 134 stars, AGPL-3.0, last commit August 22, 2026, three tools (fred_browse, fred_search, fred_get_series). It needs a FRED API key and runs from a local build or Docker:
{ "mcpServers": { "fred-mcp": { "command": "docker", "args": ["run", "-i", "--rm", "-e", "FRED_API_KEY=<your-key-here>", "stefanoamorelli/fred-mcp-server:latest"] } }}The FRED API requires a key on every request and allows 120 requests a minute before it returns 429; the server throttles itself to match. Values can be revised between pulls, which is why the official connector returns realtime_start and realtime_end. Keys, limits and vintages are in the FRED API in Python guide.
Crunchbase MCP server
Crunchbase runs its own remote server at https://mcp.crunchbase.com. It is in Claude’s connector directory (made by Crunchbase, added August 2026) and in ChatGPT’s plugin directory, and Crunchbase says it works on the free Claude plan. You sign in with your Crunchbase account through OAuth, and requests draw data tokens from your account’s allowance; when the balance reaches zero, Crunchbase blocks further requests until it resets. In Claude Code:
claude mcp add --transport http --scope user crunchbase https://mcp.crunchbase.comThen run /mcp, select crunchbase and finish the browser sign-in. In Cursor, add {"crunchbase": {"url": "https://mcp.crunchbase.com"}} under mcpServers. Crunchbase also publishes ready-made workflows (market maps, funding screens, company tearsheets) as agent skills in crunchbase/crunchbase-mcp-skills; they need the MCP connection to work. Community servers such as Cyreslab-AI/crunchbase-mcp-server (17 stars) still need a paid Crunchbase API key, so the official server is the better route.
PitchBook MCP server and the PitchBook Claude connector
PitchBook’s MCP server is the PitchBook Premium connector at https://premium.mcp.pitchbook.com/mcp, made by PitchBook and listed in Claude’s directory since November 2025. Its 16 tools search PitchBook and read company profiles, deals, investors, financials, cap tables, funds, LP commitments and team members, plus analysis of news, call transcripts and PitchBook’s reports. PitchBook’s help center says you need “a seat-based, unlimited, or trial PitchBook license” and single sign-on credentials, and lists Claude, ChatGPT, Microsoft 365 Copilot and several others as supported platforms.
In an August 2026 investor Q&A, PitchBook’s parent Morningstar said subscribers use the premium connectors “at no additional charge today,” and that expanded access is expected to move to consumption-based pricing. There is no open-source PitchBook server worth considering: without a PitchBook licence there is no data to reach. PitchBook vs Crunchbase compares the two licences, including what each lets you load into an AI tool.
Running several research MCP servers at once
You can attach all of these to one client, and the client’s limits decide how well it goes. Claude Code defers tool definitions with tool search, loading only tool names and server instructions at the start of a session, so adding servers costs little context; it caps a single tool result at 25,000 tokens by default (MAX_MCP_OUTPUT_TOKENS raises it). With tool search off, every tool definition loads up front, and the main servers on this page expose at least 80 tools between them.
Stitching many servers together has limits that no client setting fixes:
- Overlapping tools split the routing. Two PubMed servers, or the arXiv server and a multi-source one, give the model several tools that do the same thing. The edgartools README puts it plainly: “Attach one per client, not both: they overlap and split your assistant’s routing.” Pick one server per source.
- Each server returns its own shape. A paper is a PMID in one server, a DOI in another and an OpenAlex ID in a third. OpenAlex’s
get_workand the PubMed servers’ ID converters can join them, but the model has to do the joining, and it can get it wrong. - Rate limits are per source and per IP, not per server. arXiv’s limit covers every machine you control, and NCBI counts all requests from your IP address, so two servers calling the same API share one budget.
- Keys and sign-ins multiply. FRED (open source), NCBI, Semantic Scholar and edgartools each want a key or identity in their own environment variable; OpenAlex, FRED (official), Crunchbase and PitchBook each want their own OAuth sign-in.
- Everything a server returns is untrusted text. The arXiv server’s security policy warns that paper text, abstracts and LaTeX can carry prompt-injection text aimed at your client. The more servers you attach, especially ones that can write or download, the more that matters.
- Nothing checks across sources. Each server returns what its source says. Whether a paper’s claim matches the trial registry, or a 10-K figure matches the press coverage, is left to the model. If that matters, check the citations in the output.
One MCP server that covers many research sources
If you would rather attach one server than ten, three options cover several of these sources at once.
- openags/paper-search-mcp (3,097 stars, MIT, last commit October 2, 2026) searches arXiv, PubMed, bioRxiv, medRxiv, Google Scholar, Semantic Scholar, Crossref, OpenAlex, PubMed Central, Europe PMC, CORE and 11 others in one call, and deduplicates the results. It also has an optional Sci-Hub download path, off by default; the README notes that legal risk varies by jurisdiction.
- Claude Science ships built-in, read-only connectors to public research databases, mostly in the life sciences, including the Literature Graph (OpenAlex and arXiv) and FDA drug data. They need no account, except that searching OpenAlex through the Literature Graph needs a free OpenAlex API key.
- Hyperresearch, the AI deep research API, runs the Answers MCP server, which reads 120+ structured public sources, among them SEC EDGAR, FRED, ClinicalTrials.gov, PubMed, OpenAlex and arXiv, and checks each answer word for word against the quotes it cites. It does not read Crunchbase or PitchBook; the full list is on the sources page.
