Managed service · Private beta
JSON API + MCP
Ask more. Scan less.
Let your teams and their AI assistants explore your data at a fraction of the scan cost. Get answers immediately, and refine when more precision matters.
Explore 5 ready-to-query datasets from Claude, ChatGPT, Copilot, or your own agent.
Why Hilbium
More questions. Less work to make them possible.
For data teams
New questions. No custom pipelines.
Connect your tables; Hilbium prepares each one. No indexes to choose. Ask new questions without building any additional pipeline.
For your organization
Bring data exploration to more teams.
Let employees explore your data from the AI assistant they already use, with controlled read and spending budgets.
For agent developers
Give agents a budget, and room to explore.
Start at 0% source-data reads. Let the agent inspect the answer’s uncertainty and request more precision when it needs it. Keep exploration under an explicit read budget.
For shared dataPlanned
Share your data. Decide how much it reveals.
Open a dataset to partners, customers or other teams. They get aggregated answers with bounds, never rows, and you set how much of the source data their questions can read. Built into the read budget every query already carries.
You set how much it reads. Hilbium chooses what.
Average fare by passenger count, across 41 million NYC taxi trips.
NYC Taxi, yellow trips 2024 · Recorded benchmark
- Each bar is the width of a group's bounds; the dashed outline is that width at 0%.
- The caret is the true average, computed on every trip of the source. Every bound contains it.
Methodology
- Query:
SELECT passenger_count, AVG(total_amount) FROM trips WHERE passenger_count BETWEEN 1 AND 3 GROUP BY passenger_count ORDER BY passenger_count, on 41,169,720 rows. - The read budget is a ceiling. The share read is what the service reported reading for that query.
- Estimates and bounds are in dollars, as returned by the API. Latency is the median server time over nine runs, measured on 2026-11-06.
- Raising the budget step by step from 0% gave the same bounds and the same share read as each query run directly at that budget.
- Trips per group, counted on the source data: 28,632,703 with 1 passenger, 5,410,774 with 2 passengers, 1,282,073 with 3 passengers.
- True averages, computed in DuckDB on the source Parquet: $27.42, $31.20, $30.67. Every bound, at every budget, contains them.
- Four budgets recorded: 0%, 1%, 5% and 25%, shown without interpolation.
- A recorded run, not a live query.
How it works
Give your AI room to explore.
AI assistants and agents ask follow-up questions, test hypotheses, and investigate unexpected results. Hilbium lets them control how much data each question is worth reading.
01
Ask
Start with a small read budget, or none at all.
02
Inspect
Get the answer and how uncertain it is.
03
Refine
Raise the budget when it matters. Hilbium reads what narrows the bounds most.
04
Explore
Move on to the next question.
At a 0% budget, Hilbium answers from statistics prepared at ingestion, without reading a source row.
JSON API
Send a query with its read budget. Refine an answer by its identifier.
MCP
The same operations as tools your agent can call: list datasets, describe columns, query, improve an answer.
Chat assistants
Add Hilbium to Claude, ChatGPT, Copilot, or any other assistant that supports MCP. Then ask in plain language.
Benchmarks
Measured on real datasets.
16
ClickBench queries answered without reading a source row.
≤ 2 ms
for each of the six exact answers, over 99,997,497 rows.
10 of 10
most visited URLs found without reading a row, each count within its bounds.
Deployment
A managed service. Nothing to install.
Available today
Managed by Hilbium
We host and operate the service. We ingest a copy of your data from your lake; your agents and your team query it through the JSON API or MCP.
Planned
The same service, deployed inside your own cloud account.
Private beta · Access on request
Five datasets. 4.45 billion rows. Ready to explore.
From Ethereum transactions to scientific publications and NYC taxi trips. Explore with Claude, ChatGPT, Copilot, or your own agent, ask your own questions, and see how much scanning you save.
Ethereum Transactions
3,774,165,050 rows
Ethereum blockchain transactions · 21 fields
“Which address receives the most transactions?”
0xdac1…1ec7, the USDT contract · 375.7M – 377.4M transactions · 19 ms
OpenAlex
476,951,866 rows
Scientific publications and metadata · 124 fields
“How many works have been retracted?”
100,133 · Exact · < 1 ms
ClickBench
99,997,497 rows
Web analytics events · 105 fields
“How many hits came through an ad engine?”
630,500 · Exact · < 1 ms
TPC-H Lineitem · SF10
59,986,052 rows
Synthetic order line items · 16 fields
“How many line items were returned?”
14,808,183 · Exact · < 1 ms
NYC Taxis · 2024
41,169,720 rows
New York City yellow taxi trips · 19 fields
“How many trips started at JFK?”
1,989,994 · Exact · 5 ms
Each answer came back at a 0% read budget, through the MCP server, on 2026-11-06. Times are median server times over fifteen runs.
1Request access
Each request is reviewed.
2Connect
Add Hilbium to Claude, ChatGPT, Copilot or your agent with the details you receive, and sign in.
3Ask
Start with the suggested questions, then your own.
Have your own dataset? Let’s benchmark your workload.
Tell us what your data is and the questions your teams and AI assistants ask: contact@hilbium.com