data layer for research agents.
Evidence, not links. Exactly what the source says, and exactly where it says it.
One research question.
Who is buying the world's gold?
the web and specialist datasources.
One question, in the languages it needs.
A source answers in its own words, not in yours. So one search asks in many wordings at once.
- Who is buying the world's gold?
- Kdo kupuje světové zlato?
- Kto kupuje światowe złoto?
search and ranking.
Put the source itself first.
Each result says who stands behind it, so your agent can rank the institution above the story about it. Copies count once. What was left out is said.
contents and highlights.
Not only the page. The lines that answer.
In the source's own words, each with the place it stands: its section, its page, its cell in a table.
reasoning and drafting.
Every claim knows what it rests on.
Your agent reasons and drafts. Nothing it writes loses its way back to the lines.
verification.
Checked, not believed.
A person in the loop, with a better seat.
the report.
Every sentence leads back to its lines.
Cite the sentence, not the page.
Who is buying the world's gold?
5 official sources were asked. 4 could be read.
2 are buying and 2 are holding.
The Czech National Bank is buying, and says so: about 12 tonnes at the end of 2022, 90 at the end of October 2026.
It aims at 100 tonnes by 2028. At the end of 2019 it had 8, the least in its history.
China bought in every month of 2026. Its gold went from 7419 in January to 7830 in October, in units of 10,000 ounces.
Switzerland is not buying. Its 1,040 tonnes have not changed since the end of 2008.
Japan is holding: 27.20 million troy ounces at the end of December 2025, and 27.20 at the end of October 2026.
A value is not a weight. Switzerland's gold rose by 36,303.4 million francs in 2025, and all of it is price.
China's gold fell from 3695.82 to 3064.11, in units of 100 million dollars, between January and October 2026, while it bought in every month.
Point at a sentence.
Its passage opens, as the source has it.
Zlato ČNB
Obnova zlatého pokladu
Zatímco na konci roku 2022 držela přibližně jen 12 tun tohoto cenného kovu, na konci října 2026 už měla v devizových rezervách 90 tun zlata.
71.6 tonnes
While it held only about 12 tonnes of this precious metal at the end of 2022, by the end of October 2026 the total had reached 90 tonnes of gold in its foreign exchange reserves.
Czech National Bank, read on 16 November 2026.
cnb.cz/cs/financni-trhy/zlato-cnb/
cnb.cz/en/financial-markets/the-cnbs-gold/
Zlato ČNB
Historie zlatého pokladu
Nejméně v historii, 8 tun, měla centrální banka na konci roku 2019.
Obnova zlatého pokladu
Cílem ČNB je do roku 2028 zvýšit zásoby zlata na 100 tun, což by bylo nejvíce v historii centrální banky.
Czech National Bank, read on 16 November 2026.
cnb.cz/cs/financni-trhy/zlato-cnb/
官方储备资产(2026)
4. 黄金 Gold
China, State Administration of Foreign Exchange, read on 16 November 2026.
safe.gov.cn/safe/2026/0206/27116.html
Questions and answers on the SNB's balance sheet
How much gold does the SNB hold?
The SNB's gold holdings amount to 1,040 tonnes and have been unchanged since the end of 2008.
Swiss National Bank, read on 16 November 2026.
snb.ch/en/services-events/digital-services/faq-overview/
外貨準備等の状況
(令和7年12月末現在)
(重量[百万トロイオンス])
外貨準備等の状況
(令和8年10月末現在)
(重量[百万トロイオンス])
Japan, Ministry of Finance, read on 16 November 2026.
mof.go.jp/policy/international_policy/reference/
official_reserve_assets/data/0712.html, 0810.html
The SNB's annual financial statements and profit
Balance sheet as at 31 December 2025
In CHF millions
Switzerland raised its gold by 46 percent in 2025.
The SNB's gold holdings amount to 1,040 tonnes and have been unchanged since the end of 2008.
Swiss National Bank, read on 16 November 2026.
snb.ch/en/the-snb/organisation/annual-report-profit
官方储备资产(2026)
China cut its gold by 17 percent in 2026.
China, State Administration of Foreign Exchange, read on 16 November 2026.
safe.gov.cn/safe/2026/0206/27116.html
The site turned a program away.
614 tonnes
Poland, Narodowy Bank Polski
nbp.pl/zloto-nbp/
Research goes wrong at the source.
Menus glued to headlines, an ad inside a paragraph, a crossed-out line served as fact. Then the answer is wrong, and it sounds right.
Tables stay tables. Struck text stays struck.
Where an overconfident wrong answer is unacceptable.
Legal
Case citations checked before filing
A citation to a case that does not exist is found by your agent, not by the court.
Finance
Figures from filings, each traced to its cell
Revenue, margin, and debt as figures that name their table, their row, and their column.
AI products
RAG that cites the passage, not the page
Each chunk is one whole paragraph or one whole table, and each answer cites its passage.
AI labs
Reinforcement learning for research agents
An environment that stays the same: the same text in every rollout, time fixed on one day.
Science and medicine
Systematic reviews: the search writes its own PRISMA record
For each datasource the date, the wordings, and the counts, written up in PRISMA form.
Compliance
Contract review that tells a missing clause from an unread page
A clause that is not found says whether the contract lacks it or a page could not be read.
From 'trust me' to 'check it'.
Of the sentences in a report,
how many lead back to their exact lines.
95.6%
Measure it yourself. Compare it yourself.
- getclean. 95.6%
- exa 67.4%
- firecrawl 63.8%
- tavily 55.2%
Connect your research agent to the data layer.
You pay when a source is read.
Reading it again is free.
pip install getclean npm install getclean ai.getclean:getclean go get github.com/deterministic-labs/getclean-go dotnet add package GetClean npx -y getclean-mcp from getclean import GetClean
client = GetClean()
record = client.fetch_contents(
url="https://www.cnb.cz/en/financial-markets/the-cnbs-gold/",
highlights=["How much gold does the bank hold?"],
)
print(record.highlights[0].text)import { GetClean } from "getclean";
const client = new GetClean();
const record = await client.fetchContents({
url: "https://www.cnb.cz/en/financial-markets/the-cnbs-gold/",
highlights: ["How much gold does the bank hold?"],
});
console.log(record.highlights?.[0]?.text);GetClean client = GetClean.builder().build();
Contents record = client.fetchContents(
FetchContentsRequest.builder()
.url("https://www.cnb.cz/en/financial-markets/the-cnbs-gold/")
.build());
System.out.println(record.getOutcome().getKind());c := getclean.NewClient()
record, err := c.FetchContents(ctx,
&getclean.FetchContentsRequest{
URL: "https://www.cnb.cz/en/financial-markets/the-cnbs-gold/",
})
if err != nil { return err }
fmt.Println(record.Outcome.Kind)var client = new GetCleanClient();
var record = await client.FetchContentsAsync(
new FetchContentsRequest
{
Url = "https://www.cnb.cz/en/financial-markets/the-cnbs-gold/"
});
Console.WriteLine(record.Outcome.Kind);A key under which no model works.
A key that fetches only the sites you name.
Nothing kept after the response, if you say so.
1,000 research runs
One question each, 6 sources read. An example.
- 1,000 searches
- $10
- 5,000 wordings of the question
- $5
- 2,000 steps by a model
- $4
- 6,000 pages read
- $30
- 1,000 pages that could not be read
- free
- 14,000 quotes found word for word
- free
- 2,000 pages compared, cell by cell
- free
- In all
- $49
start
$0 a month
Pay as your agent reads
at the prices of the bill.
A spending limit on every call.
team
$99 a month
2,500 research runs
a month are in it.
A key for each agent, each with its own budget.
scale
$499 a month
15,000 research runs
a month are in it.
Higher rate limits, and more calls at once.
enterprise
custom
As many runs as you need
by contract.
An invoice, and the people who built it to ask.