web-researcher-mcp
Your AI research assistant. It searches, reads what you can open, and checks its sources.
Real search and primary sources. The pages behind your own logins, like LinkedIn and X. Then it checks every citation, keeps your research, and runs on your machine. No API key needed to start.
43 seconds on one part of it, checking citations. A cartoon of the problem, then real tool runs: a made-up case comes back not found, and a reference list is checked for fake DOIs, retractions and dead links.
A research desk your AI can use
It finds the sources, reads them as you would, checks them, and keeps the work. All on your machine.
Search everywhere you'd look
Web search you can limit to sites you trust (lenses), plus SEC filings, US court records, patents, clinical trials, economic data and full-text papers. It works with no API key.
Work as you
Opt-in saved logins let it read pages you can already open: LinkedIn, X, Facebook, Instagram, YouTube and more. You sign in yourself. The login is encrypted and yours alone.
Check before you rely
Dedicated tools check that a source exists, was not retracted, still resolves, and says what is claimed. They return evidence, not a verdict, and their accuracy is measured.
Read the whole page
Web pages, PDFs, Word and PowerPoint files, YouTube transcripts and forum threads, in full. Not a snippet and not a summary of a summary.
Keep it and watch it
Research sessions survive context loss. Compare answers from several models. Export citations as APA, MLA, BibTeX, RIS or CSL-JSON. Opt-in monitors flag new results, memory keeps findings across chats, and team workspaces share them.
Yours to run
Free, open source, no account. Your queries go only to the search services you pick. Run it on your laptop, or as a shared server for a team. Anything that stores personal data is off until you turn it on.
It can work as you, safely
A lot of the best information sits behind a sign-in: your LinkedIn network, an X thread, a forum, a site you pay for. With a saved login, your AI reads those pages for you. You don't paste anything in.
You sign in, once
When a page needs a login, your AI asks. You sign in yourself, in your own browser or in a private login window. 2FA and CAPTCHA work as normal. Your password is never saved.
It reads as you
From then on it reads pages on that site automatically, as you would see them. The login goes only to that site, and pages read this way never enter a shared cache.
You stay in charge
Logins are encrypted, tied to your user, and expire on their own (cookies after 24 hours and browser profiles after 7 days, by default). site_logout deletes one at any time.
Some sites, such as LinkedIn, forbid automated access in their terms. Use saved logins only where yours allow. It does not bypass paywalls, and it never signs in for you. Pages you read this way are treated as untrusted: your AI is told not to follow instructions hidden in them.
AI search tools make things up. That's a problem when your name is on the work.
Independent studies keep finding the same problem, no matter who tests it or how. Columbia Journalism Review tested eight AI search tools on real citations and found more than 60% wrong. A BBC/EBU study across 22 public broadcasters in 18 countries found 45% of AI news answers had a significant issue, most often bad sourcing. A 2026 study of Google's AI Overviews, reported by The New York Times, found only 39% were both correct and fully backed by the sources they cited. They link to papers that don't exist, invent source records, and present low-quality content with the same confidence as expert-reviewed research.
The cost is now public. The Charlotin database lists over 2,200 court decisions that involve AI-hallucinated content. From January 1, 2027, California SB 574 requires attorneys to personally verify every citation in a court filing.
If your work gets cited, published, submitted to a court, or shown to a client, "probably real" isn't good enough.
- Lawyers: "If I cite a case that doesn't exist, I get fined $50,000"
- Medical researchers: "Clinical decisions based on a health blog could hurt someone"
- Journalists: "I need to cross-check court filings, not AI summaries"
- Graduate students: "I spent 3 hours tracking down a citation my AI invented"
A real fake case, checked.
"Is this case real? Varghese v. China Southern Airlines, 925 F.3d 1339"
Varghese v. China Southern Airlines Co., 925 F.3d 1339 (11th Cir. 2009)
No such case exists. The lawyers filed it anyway and the court sanctioned them in Mata v. Avianca.
No court record for this case.
The closest hit is a 2017 conference paper about the airline. That is not a court opinion, so it is not a match. Run on 2026-10-25.
Sources: Mata v. Avianca, S.D.N.Y. 2023; Columbia Journalism Review, 2025; BBC/EBU, 2025; The New York Times, 2026.
See it work on real runs
Three actual tool runs, not mock-ups. Each one tells you what it found and how sure it is, so you decide.
Catch a made-up case
The tool also shows the nearest thing it found, so you can see why it said no.
Audit a whole reference list
DOI not found in Crossref
retraction notice 2010-02-06
One pass flags the fake DOI, the retraction and the dead link. It marks what it could not check as unchecked, not fake.
Search only trusted sites
4 results, restricted to legal and court sources
supreme.justia.com/cases/federal/us/384/436
uscourts.gov/about-federal-courts/... source tier: high
oyez.org/cases/1965/759
supreme.justia.com/cases/federal/us/560/370
A search lens keeps results on court and legal sites. Every link is a live page you can open.
Real tool output from runs on 2026-10-25, trimmed and reformatted for display. Results change over time, and every result is evidence for you to judge.
Measured, not just claimed
We test the citation checks against a labeled set of real, invented and retracted papers. These are the numbers from the latest published run.
Run on 2026-08-07 against a small, curated set, and re-run weekly in CI. Other citations will vary. The weakest check is "does the source say what is claimed": it under-flags on purpose so it never falsely accuses. Full results and how to reproduce them.
What you can ask
Plain requests. Your AI picks the tools.
Lawyers
"Check every case in this brief. Which ones can't be found, and which were overruled or retracted?"
Uses audit_bibliography verify_citation legal_search
Researchers
"Find the full text of recent papers on this topic and tell me if any were retracted."
Uses academic_search paper_fulltext verify_citation
Journalists
"Pull this company's latest 10-K risk factors and any court cases that name it."
Uses filing_search legal_search scrape_page
Students
"Put these sources in APA and tell me which links are dead."
Uses format_bibliography audit_bibliography
Market research
"Read this competitor's company page and recent posts on LinkedIn and X, then pull their latest SEC filing."
Uses site_login scrape_page filing_search
Teams
"Save this search. When I ask again, show me only what's new, and share the best finds with the team."
Uses monitor_query_save monitor_query_check workspace_contribute
Get started in 30 seconds
Try this first: ask your AI "Is Varghese v. China Southern Airlines, 925 F.3d 1339 a real case?" No sign-up is needed. It works right away using DuckDuckGo. You can connect a search service like Brave or Serper later for richer results. Works with Claude, Cursor, VS Code, Codex, Gemini CLI, and any MCP client.
Windows, Homebrew, Docker, Go install, or a different AI client? See the full documentation for every install path and client config.
How it compares
Every AI chat or agent — Perplexity, ChatGPT, Gemini, Grok, whatever you use — ships with some form of built-in web search. web-researcher-mcp isn't a rival chatbot; it's a toolset you attach to whichever one you already use, including as an MCP connector inside ChatGPT and Gemini. So this isn't a vendor-vs-vendor comparison — it's what you get by attaching it versus relying on whatever search is already built in.
| web-researcher-mcp | Your AI's built-in search | |
|---|---|---|
| You choose which search engines, sites, libraries and APIs get searched | Yes — built-in providers + custom lenses for trusted sites, academic libraries, and structured APIs | No — fixed to whatever index the vendor built in |
| Wrong citations, in independent testing | Every link comes from a live search, then dedicated tools check it exists, hasn't been retracted, and still resolves — before it ever reaches you |
Google AI Overviews — only 39% correct and fully sourced Perplexity — 37% wrong ChatGPT Search — 67% wrong Gemini — 50%+ broken or fake links Grok 3 — 94% wrong BBC/EBU study — 45% had a significant issue |
| Purpose-built tools for filings, case law, patents, clinical trials, and economic data | Yes — dedicated tools reading SEC EDGAR, CourtListener, patent offices, ClinicalTrials.gov, and statistical agencies directly | General web search only |
| Pages behind your own login | Yes, opt-in — save a login once, and scrape_page reads as you. Encrypted and tied to your user |
Public pages only |
| Multi-step research you can audit and export | Yes — session history, bibliography checks, exportable citation trail | A chat transcript, not a research record |
| Same toolset across every AI you use | Yes — identical tools and behavior in Claude, Cursor, VS Code, or as a connector in ChatGPT/Gemini | Locked to whichever app you're using |
| Who controls your search cache and history | You do — open source, runs on your machine, you set what's kept and for how long | The vendor does, on its own servers |
| You can inspect or self-host the code that's searching for you | Yes — MIT-licensed and open source, audit it or run your own instance | No — closed source |
Citation-accuracy figures: The New York Times, 2026; Columbia Journalism Review, 2025; BBC/EBU, 2025 — each testing a product's own search independently. Every other row reflects each product's own publicly documented features — the pattern holds across vendors, not just the ones named above. Your underlying AI model and search provider still see your queries either way — this table is about who gives you a way to check, control, and keep a record of what came back, not about who "sees" your data.
Building search into a product instead?
Calling Tavily, Exa, or another search API directly gets you results. It doesn't get you what a team needs to run that in production — this is what you get by going through web-researcher-mcp instead.
No single point of failure
Wire in one provider directly and its outage or rate limit is your outage. web-researcher-mcp routes across every provider you configure, with per-provider circuit breakers and automatic fallback, so one provider going down doesn't take your product's search down with it.
One integration for every domain
Web search is one API. SEC filings, court records, patents, clinical trials, and economic data are each their own, with their own quirks. Integrate once through a consistent tool contract instead of building and maintaining a separate integration for every domain you need to cover.
Multi-tenant governance, not bolted on after
Per-tenant rate limits, audit logging, and consent-gating for personal data are things you'd otherwise build yourself on top of a raw search API. web-researcher-mcp ships them, so your team isn't re-building infrastructure a shared research tool should already have.
37 tools, organized by outcome
Not a search box — a full research toolkit your AI can call on directly.
Catch fake citations
verify_citationDoes this source exist, was it retracted, does it say what is claimed?audit_bibliographyThe same checks on a whole reference list.verify_recommendationSpots self-promotion and missing outside backing in "best of" lists.archive_sourceSaves a Wayback Machine snapshot of a page (writes to an outside service).
Search everything
web_searchAny provider you set up, or only trusted sites with a lens.scrape_pageReads a page, PDF, Office file or YouTube transcript.search_and_scrapeSearches and reads the top pages in one step.news_searchRecent news with date filters.image_searchFinds images.
Academic & scientific
academic_searchPapers from scholarly sources.paper_fulltextFull text of a paper from a DOI or link.citation_graphWhat cites a paper, and what it cites.clinical_searchStudies from ClinicalTrials.gov.monarch_searchRanks diseases and genes by phenotype.
Legal, filings & markets
legal_searchUS court opinions.filing_searchSEC filings: 10-K, 10-Q, 8-K and more.patent_searchPatents.econ_searchFRED, World Bank, OECD and other economic data.
Deep, multi-step research
sequential_searchMulti-step research that keeps its sources.get_research_sessionPicks a session back up after context loss.research_exportExports a session as a report.research_panelAsks several models and shows where they disagree.
Formatting & recon
format_bibliographyAPA, MLA, BibTeX, RIS and CSL-JSON.brand_researchA company's public brand identity: colors, logos, type.company_reconCertificate and Wayback records for a domain.awesome_list_searchCommunity "awesome" lists on a GitHub topic.
Opt-in, off by default
site_loginSign in once, by hand, so pages you can open can be read as you.site_logoutremoves the login.monitor_query_saveSave a search and get told when it finds something new.monitor_query_checkruns it.memory_saveKeep findings across sessions.memory_recallbrings them back.workspace_contributeShare findings with a team.workspace_readreads them.get_my_analyticsSee your own usage.
Every tool listed here is checked against the docs automatically before each release. Full details: Tools reference.
Frequently asked questions
What is web-researcher-mcp?
A free, open-source add-on that gives your AI assistant real web search, full-page reading, and multi-source research — academic papers, patents, company financial filings, court cases, clinical trials, economic data, and more. Every result comes from a live search, not the AI's memory, and dedicated tools check that each citation exists, was not retracted, and still resolves.
How is this different from Perplexity or ChatGPT search?
Perplexity and ChatGPT, like every AI chat or agent's built-in search, hand back whatever a fixed ranking algorithm turns up and hope it's right — with no way for you to choose the sources or verify the citation afterward. It's not particular to those two: in the independent tests cited above, Perplexity got 37% of citations wrong, ChatGPT Search got 67% wrong, Grok 3 got 94% wrong, and only 39% of Google AI Overviews were both correct and fully backed by their cited sources. The pattern holds across vendors. web-researcher-mcp lets you restrict search to sources you trust using search lenses, and ships tools that check a citation's existence, retraction status, and link liveness before you rely on it.
Do I need to sign up for anything to use it?
No. It works right away using DuckDuckGo, which needs no account. But DuckDuckGo is just the zero-config default, not a ceiling — under the hood it supports a dozen search providers (Brave, Serper, Tavily, Exa, SearXNG, and more), and you can configure several at once so it automatically falls back to another if one is down or rate-limited. You can also point it at self-hosted or free-tier providers, restrict it to curated "search lenses" for a field like law or medicine, or route straight to structured sources like SEC EDGAR, CourtListener, or patent offices instead of general web search. Connecting any of that just means signing up with the provider and pasting in the key they give you — nothing is required to get started, and everything past DuckDuckGo is optional.
Is my research data private?
Yes. The server runs on your own computer. Your search queries go directly from your computer to whichever search service you've chosen — there's no middleman server collecting or seeing what you research. No account is created and nothing is tracked. Features that store personal data (saved logins, memory, analytics, shared workspaces) are off by default, need your consent, and are encrypted when on.
What is a search lens?
A curated list of trusted sites for a field — medical journals, court databases, government filings, academic repositories. Using a lens restricts your AI's search to only those sites, so results come from sources you already trust instead of the open web.
Does it work with Claude, Cursor, and other AI tools?
Yes. It's built on the Model Context Protocol (MCP), so it works with Claude Code, Claude Desktop, Cursor, VS Code, LM Studio, Cline, and any other assistant that supports MCP tool use.
Is it really free?
Yes — free and open source under the MIT license, forever. There's no paid tier and no usage limits from the project itself. The only cost you might incur is if you choose to sign up for a paid search service for better results, which is entirely optional.
Does it check that citations are real?
Yes. verify_citation checks that a source exists, matches a real record, was not retracted, and still resolves. audit_bibliography does this for a whole reference list in one pass, and also flags entries whose source does not say what the citation claims. The tools return evidence, not a verdict, so you decide what to believe. Their accuracy is measured on a labeled gold set that is published in the docs.
Can it read pages that need a login, like LinkedIn?
Yes, if you turn it on. site_login saves a login you create yourself, either from your own browser or from a private login window where you sign in by hand (2FA works). scrape_page then reads pages behind it as you. We have tested LinkedIn, X, Facebook, Instagram and YouTube. It is off by default, needs your consent, and the saved login is encrypted, tied to your user, and removed by site_logout. Your password is never saved. It does not bypass paywalls. Some sites, such as LinkedIn, forbid automated access in their terms, so use it only where yours allow.
How many tools does it have, and what can it search?
37 tools covering web, image, and news search; academic papers with full text; patents; company financial filings; court cases; economic data; clinical trials; full-page and document reading; citation checking and bibliography auditing; multi-model research panels; and multi-step research sessions you can export. A few opt-in tools (saved logins, saved-query monitoring, memory, shared workspaces, personal analytics) stay off until you turn them on.