1. Why AI answers need grounding
A language model writes the most plausible continuation of your question. For explaining a concept that works well. For research it fails in a specific, costly way: the model produces citations to papers that do not exist, statistics with no dataset behind them, and case numbers that look right but lead nowhere. It is not lying; it has no record to consult, so it reconstructs one.
Grounding means the answer is written from material retrieved for this question, and every claim points back to that material. Uploading PDFs grounds an answer in documents you already have. Connecting databases grounds it in the records you do not have yet: the trial registry, the statistics table, the archive volume, the consolidated regulation.
2. What an MCP server is, in plain terms
The Model Context Protocol is an open standard, introduced in late 2024 and now supported across the major AI platforms, for connecting AI assistants to outside tools and data. Think of it as a universal adapter: instead of every AI product writing its own integration for every database, a database gets one MCP server and any compatible assistant can use it.
An MCP server does three things:
- Describes its tools, for example “search trials by condition and phase” or “fetch a legal act by CELEX number”, with the parameters each tool accepts.
- Runs the call against the underlying database when the assistant asks.
- Returns structured results with the database’s own identifiers intact.
Some servers are run by the institution behind the data, such as the Swedish National Archives, the Wikimedia Foundation and the Open Targets consortium. Others are open-source community servers wrapping a public API.
3. What happens when you ask a question
- 1
You ask
A question in plain language, in any language you work in.
- 2
The assistant picks a tool
It reads what your connected sources can do and chooses, say, a dataset search.
- 3
The server queries the database
The MCP server translates the call into the database’s own API or query language.
- 4
Records come back with IDs
Dataset codes, NCT numbers, CELEX numbers, DOIs, archive reference codes.
- 5
The answer cites them
Every claim points to a record you can open and check in the original.
The important part is step four. Identifiers survive the round trip, so the finished answer is not a summary you have to trust. It is a set of references you can open.
4. MCP compared with web search, uploads and model memory
These approaches are not rivals. Each grounds a different kind of question, and a good research chat uses all of them.
| Property | Model memory | Web search | Your uploads | Database via MCP |
|---|---|---|---|---|
| Up to date | No, frozen at training | Yes | As current as your files | Yes, queried live |
| Exact identifiers | Often invented | URLs of mixed quality | Your own documents | Native IDs: DOI, NCT, CELEX |
| Precise filters | No | Keywords only | No | Years, regions, units, phases |
| Primary records | No | Rarely | Only what you have | Yes, the database itself |
| Best for | Explaining concepts | News and grey literature | Close reading of known texts | Facts, figures and records you must cite |
5. Six worked examples by field
Each example uses a source you can connect in Kahubi today. The prompts are written to be copied; the steps describe what the assistant does with the database’s tools.
“Search the Swedish National Archives for court records from Dalarna in the 1670s that mention witchcraft (trolldom). List what you find with reference codes, and tell me where the transcription looks unreliable.”
- 1.Runs a full-text search over handwriting-recognised pages, trying period spelling variants
- 2.Opens the most relevant volumes to read the surrounding pages
- 3.Returns each passage with its archive reference code and a link to the page image
You get: Passages you can open in the original scans, with transcription doubts called out instead of smoothed over.
“Using Eurostat, compare youth unemployment (ages 15 to 24) in Sweden, Finland and Spain from 2015 to 2024. Name the dataset code and unit, and flag any breaks in the series.”
- 1.Searches Eurostat for the matching dataset
- 2.Inspects its dimensions: age band, unit, sex, geography and time
- 3.Queries only the rows you asked for
You get: A table where every number traces back to a dataset code and unit, not a figure half-remembered from a news story.
“Find completed phase 3 trials of GLP-1 receptor agonists in heart failure with posted results on ClinicalTrials.gov, then find the PubMed articles that report them.”
- 1.Filters the registry by condition, intervention, phase and status
- 2.Reads the posted outcome measures for each study
- 3.Searches PubMed for publications linked to each NCT number
You get: Trials listed by NCT number next to the PMIDs of their publications, so registered and published outcomes can be compared.
“Which targets have the strongest genetic evidence for Parkinson’s disease, and which already have drugs in clinical development? Give Ensembl and UniProt identifiers.”
- 1.Resolves the disease to its ontology ID in Open Targets
- 2.Queries association scores broken down by evidence type
- 3.Maps each target to its UniProt accession through TogoMCP
You get: A ranked target list with evidence types, drug phases and identifiers, ready for your own follow-up analysis.
“Which EU acts amend or are based on the GDPR, and which Court of Justice judgments interpret the right to erasure in Article 17? Cite CELEX numbers.”
- 1.Resolves the GDPR to its CELEX number
- 2.Follows the amendment and legal-basis relationships between acts
- 3.Searches case law that interprets the article
You get: Acts and judgments listed by CELEX number with dates, so every citation opens the official text.
“How did European news coverage of heat waves change in volume and tone over the last three months? Break it down by country and give example articles.”
- 1.Builds a coverage timeline for the topic
- 2.Computes tone and a breakdown by source country
- 3.Pulls sample articles from the peaks
You get: A timeline with the articles behind each spike, and a clear note that GDELT article search reaches back three months.
Run these on your own research questions
80 databases, one click each, cited answers in the same chat as your library. Free plan included.
6. How to write prompts that stay grounded
The same database gives very different answers depending on how you ask. Compare:
“What does research say about youth unemployment in the Nordics?”
No source, no years, no measure. The assistant may answer from memory and you cannot tell which numbers are real.
“Using Eurostat, give youth unemployment (ages 15 to 24, percent of the labour force) for Sweden, Norway, Denmark and Finland, 2015 to 2024. Name the dataset code. If a country or year is missing, say so instead of estimating.”
Source, measure, unit, scope and a rule for gaps. Every number can be traced.
- Name the source. “Using Eurostat” or “search the Swedish National Archives” steers the assistant to the right tool instead of a general web search.
- Ask for identifiers. Request DOIs, NCT numbers, CELEX numbers, dataset codes or reference codes in the answer. Identifiers make every claim checkable in seconds.
- Set the scope up front. Years, regions, age bands, trial phases and units belong in the question, because databases filter precisely and models guess.
- Invite the empty answer. “If nothing matches, say so” keeps the assistant from filling gaps with plausible-sounding text.
- Retrieve first, interpret second. Ask for the records, skim them, then ask for the synthesis. You catch a wrong dataset before it shapes a conclusion.
- Chain sources on purpose. Resolve names with Wikidata before searching an archive, or go from a trial registry to the publications that report it.
7. Limits worth knowing
Grounding moves the question from “is this true?” to “is this record the right one?”. That is progress, not the finish line.
- Coverage is the database’s coverage. ClinicalTrials.gov does not include trials registered only in the EU’s CTIS, and GDELT article search reaches back three months. Absence in one database is not absence in the world.
- Sources carry their own error. Handwriting recognition misreads words, preprints are not peer reviewed and statistical series change definitions. A grounded answer inherits these, so a good assistant says which apply.
- Numbers need their units. Keep the unit, reference year and dataset code next to every figure you report. The same indicator often exists in several variants.
- Servers can be slow or go offline. Especially community-run ones. Kahubi health-checks every source and stops offering a failing one to the assistant until it recovers.
- Retrieved text is data, not instructions. A record can contain any text. Kahubi marks tool results as retrieved material so nothing inside a record can redirect the assistant.
8. Set it up in Kahubi
Desktop AI apps usually ask you to edit configuration files and run MCP servers locally. Kahubi hosts the connection for you, next to your library, drafts and analyses.
- 1
Create a free account. It takes a minute and needs no credit card.
- 2
Connect your sources. Open Settings, Integrations, Data sources and click Connect. Most sources need no API key, and “Recommended for you” starts from your university’s country.
- 3
Ask in chat. Connected sources appear in the composer, where you switch them on or off per question. When a question fits a source you have not connected, Kahubi suggests it right in the conversation.
Connected sources work alongside what Kahubi already searches on its own: scholarly papers through OpenAlex, your uploaded library and the web. Running your own server? Paste any public streamable HTTP MCP endpoint under Data sources, test it and choose which tools the assistant may use.
9. Databases you can connect today
The catalog grows continuously. Every source is tested with live queries before it is listed. Open a source to see example questions and the tools it offers.
Archives & heritage
- Riksarkivet · Search AI-transcribed handwritten records from the 1600s onward, the national archive catalogue, church books, court records and more.
- Europeana · Search 50+ million digitised items from European museums, libraries and archives.
- Trove · Australian newspapers, books, images, maps and archives from the National Library of Australia.
- Libris · The Swedish national union catalogue: books, journals, theses, maps and music in Swedish libraries.
- Kubhist historical newspapers · Full-text concordance search in digitised Swedish newspapers 1645–1926, by title and decade.
- Nasjonalbiblioteket · Norwegian newspapers from 1763, books, periodicals, photos and maps with full-text search.
- Gallica · French national library: newspapers, manuscripts, maps, books and images, with OCR full-text search.
- Delpher · Dutch newspapers 1618–1995 with article-level OCR text.
- Internet Archive · Digitised books, periodicals, audio, film and software from archive.org, with full text for scanned texts.
- HathiTrust · Look up digitised library volumes by OCLC, ISBN, LCCN, ISSN or HathiTrust id.
- Open Library · Books, editions and authors from the Internet Archive’s Open Library, including search inside scanned texts.
- Smithsonian Open Access · Objects, specimens and archives from the Smithsonian museums and research centres.
- Nationalmuseum · Sweden’s national museum of art and design: paintings, sculpture, drawings, prints and applied art.
- The Met collection · The Metropolitan Museum of Art’s 490,000+ objects, most with public-domain images.
- Art Institute of Chicago · The Art Institute of Chicago collection with IIIF images.
Scholarly literature
- OpenAlex · Works, authors, institutions and citation graphs across 270M+ scholarly records.
- Crossref · DOI metadata, reference lists, journals, funders and publisher members.
- ORCID · Researcher profiles, works, affiliations, funding and peer-review records.
- arXiv · Preprint search with metadata, category browsing and full-paper reading.
- Europe PMC · Life-science literature: 40M+ abstracts, preprints and open-access full text.
- OpenAIRE Graph · Open-science graph linking publications, datasets, software and projects across Europe and beyond.
- DOAJ · Peer-reviewed open-access journals and their articles, with APC and licence details.
- bioRxiv & medRxiv · Biology and medicine preprints: search, recent postings by category, full text and the published journal version.
- HAL · 1M+ full texts from French research: articles, theses, conference papers and preprints.
- zbMATH Open · Mathematics reviews and bibliography since 1868 with MSC classification.
Medical & clinical
- PubMed & Europe PMC · Biomedical literature: search, MeSH terms, related articles, citations and lawful full text.
- ClinicalTrials.gov · Search registered clinical trials, read full study records, posted results and eligibility criteria.
- Clinical evidence · ClinicalTrials.gov, openFDA drug labels, adverse-event reports and approvals alongside PubMed.
- CDC data and WONDER · US public-health datasets (data.cdc.gov) and CDC WONDER mortality and natality queries.
- openFDA · FDA drug and device adverse events, recalls, labels and shortages.
- WHO Global Health Observatory · WHO health indicators for every member state: life expectancy, mortality, disease burden, health systems.
- USDA FoodData Central · Nutrient composition of foods from the USDA reference database.
- Open Food Facts · Crowd-sourced database of 3M+ packaged food products: ingredients, nutrition, labels, Nutri-Score.
Life science
- TogoMCP · One service across 30+ life-science databases: proteins, genes, compounds, structures, pathways, variants and MeSH.
- Open Targets Platform · Target–disease–drug associations with the evidence behind them, from EMBL-EBI.
- Protein structures · Search experimental and predicted protein structures, find similar folds, track ligands and fetch annotations.
- NCBI E-utilities · Gene, ClinVar, dbSNP, Taxonomy, Nucleotide, Protein, PMC and other NCBI databases.
- GBIF · 2.5 billion species occurrence records from museums, surveys and citizen science worldwide.
- iNaturalist · Citizen-science observations with photos, research-grade identifications and places.
- Paleobiology Database · Fossil occurrences, taxa, collections and diversity through geological time.
Official statistics
- Eurostat · EU official statistics: economy, demography, trade, health and NUTS regions.
- OECD · OECD statistics on economy, education, health, labour and environment (SDMX).
- IMF data · International Monetary Fund macroeconomic datasets.
- Statistics Sweden · Swedish official statistics: population, economy, labour market, education and environment.
- Sveriges Riksbank · Policy rate, exchange rates and interbank rates from the Swedish central bank.
- Kolada · 6,000 key indicators for every Swedish municipality and region: schools, care, finances, environment.
- Statistics Norway · Norwegian official statistics from the StatBank.
- Statistics Denmark · Danish official statistics from Danmarks Statistik.
- World Bank · Country profiles, World Development Indicators and poverty data for 200+ economies.
- ECB Data Portal · Euro-area monetary, financial and macro statistics from the European Central Bank.
- ILOSTAT · ILO labour statistics: employment, unemployment, wages, working conditions by country.
- UN Comtrade · International goods trade by reporter, partner, year and HS commodity code.
- US Census Bureau · American Community Survey, decennial census and other Census datasets by geography.
- NOAA climate data · Historical climate observations and normals from NOAA’s National Centers for Environmental Information.
- SMHI meteorological observations · Swedish weather station observations: temperature, precipitation, wind, snow, back to the 1800s for some stations.
- Open-Meteo historical weather · Daily historical weather for any coordinate since 1940 (ERA5 reanalysis), plus geocoding.
- USGS Earthquake Catalog · Global earthquake catalogue (ComCat) searchable by time, magnitude and location.
- OpenAQ air quality · Ground-level air-quality measurements from government and research monitors worldwide.
Knowledge bases
- Project Gutenberg · Search and read 70,000+ public-domain books.
- Getty Vocabularies · Controlled vocabularies for art, architecture, historical place names and artist names.
- OEIS · The On-Line Encyclopedia of Integer Sequences: identify a sequence from its terms.
- Wikidata · Structured knowledge graph: entity search, statements and SPARQL queries.
- Wikipedia · Article search, summaries, full articles and sections in any language edition.
- Glottolog · Reference catalogue of the world’s languages, dialects and families.
- MusicBrainz · Open music encyclopedia: artists, works, releases, recordings and relationships.
Datasets & code
- DataCite · DOIs for research datasets, software, samples and other non-article outputs.
- Zenodo · CERN’s open repository for datasets, software, papers, posters and supplementary material.
- Hugging Face Hub · Search models, datasets, Spaces and papers on the Hugging Face Hub.
- Stack Exchange · Q&A across Stack Overflow, Cross Validated, MathOverflow, Academia and 170+ sites.
Law & politics
- EUR-Lex · EU legislation, treaties, preparatory acts and Court of Justice case law with their amendment graph.
- Riksdagen · Bills, committee reports, motions, protocols, SOU reports and SFS statutes with full text.
- CourtListener · US court opinions, dockets, citations, judges and oral arguments from the Free Law Project.
- US Federal Register · Federal rules, proposed rules, notices and presidential documents since 1994.
- Congress.gov · US bills, laws, members, committees, votes, nominations and CRS reports.
- SEC EDGAR · US company filings: 10-K, 10-Q, 8-K, insider transactions and institutional holdings.
- CORDIS · EU-funded projects and results from FP7, Horizon 2020 and Horizon Europe.
- NSF Awards · US National Science Foundation grants by keyword, investigator or institution.
News & places
- US National Weather Service · Forecasts, alerts, observations and forecast discussions for the United States.
- GDELT · Search three months of world news in 65+ languages; coverage volume, tone and country breakdowns over time.
- OpenStreetMap · Geocode places, reverse-geocode coordinates and query map features near a point or within an area.
Last updated 2026-09-15.