DataTug is an open-source workspace where you and AI investigate connected data: in the web app, in the terminal, or inside your own coding agent with the DataTug AI skills. Answers across sources, with their evidence, work today. Shared meaning and memory are what we are building next.
The USA buys the most in total, yet ranks number 17 per person.
Ask a follow-up| Country | Sales, USD | Population (2025) | USD per million people |
|---|---|---|---|
| Ireland | 45.62 | 5,484,367 | 8.32 |
| Czech Republic | 90.24 | 10,886,878 | 8.29 |
| Finland | 41.62 | 5,646,436 | 7.37 |
| Canada | 303.96 | 41,651,653 | 7.30 |
Swipe for the other views
From the real demo run: one saved query, no model.
Work in the browser or the terminal, or let your coding agent do it. Each part shows how finished it is.
The whole platform: datatug.io
The web app: explore, query and chat over your data in the browser.
The same projects in a keyboard-driven terminal UI: run datatug.
Your coding agent uses DataTug as a tool: skills for the CLI, and an MCP server.
Interactive DataTug views inside your agent's chat.
Not a chain of thought: a record of actions, decisions and evidence. First, exactly what the real run did. Then the investigation as DataTug shows it, with the steps that still depend on features ahead marked.
The DataTug CLI executed one saved DTQL query. It joins three sources and calls no model.
Command
$ datatug query run sales/chinook-sales-per-capitaWhat it printed
query download: 412 rows chinook
query download: 24 rows geo
query download: 216 rows geo
query process: 24 rowsCross-source DTQL joins in the CLIAvailable
Reproduce it: datatug/datatug-demo-projects, commit 0e5b98f, DataTug CLI 0.51.0.
The same question as an investigation in DataTug.app. Every step is marked: it either happened in the real run, or it depends on a feature still ahead, shown with its status.
Which countries buy the most music relative to their population?
from:
database: chinook
name: Invoice
alias: i
joins:
- from:
database: geo
name: country_aliases
alias: a
joins:
- from: {database: geo, name: population_wb, alias: p}
on: [{left: {field: country, source: a}, op: '==', right: {field: country, source: p}}]
on: [{left: {field: BillingCountry, source: i}, op: '==', right: {field: alias, source: a}}]Where knowledge comes from
Steps marked "In the real run" happened. Steps marked "Target behaviour" depend on features that are still ahead, and each carries its status.
Sort it, filter it, resize it, select a bar to find the row. The grid and the chart are two views of the same result.
| Country | Sales, USD | Population (2025) | Sales per million people |
|---|---|---|---|
| Ireland | 45.62 | 5,484,367 | 8.32 |
| Czech Republic | 90.24 | 10,886,878 | 8.29 |
| Finland | 41.62 | 5,646,436 | 7.37 |
| Canada | 303.96 | 41,651,653 | 7.30 |
| Portugal | 77.24 | 10,804,871 | 7.15 |
| Norway | 39.62 | 5,610,870 | 7.06 |
| Denmark | 37.62 | 6,009,169 | 6.26 |
| Hungary | 45.62 | 9,514,251 | 4.79 |
| Austria | 42.62 | 9,208,163 | 4.63 |
| Sweden | 38.62 | 10,596,620 | 3.64 |
| Belgium | 37.62 | 11,941,781 | 3.15 |
| France | 195.10 | 68,720,337 | 2.84 |
| Chile | 46.62 | 19,859,921 | 2.35 |
| Netherlands | 40.62 | 18,087,633 | 2.25 |
| Germany | 156.48 | 83,491,249 | 1.87 |
| United Kingdom | 112.86 | 69,487,000 | 1.62 |
| USA | 523.06 | 341,784,857 | 1.53 |
| Australia | 37.62 | 27,614,411 | 1.36 |
| Poland | 37.62 | 36,435,861 | 1.03 |
| Brazil | 190.10 | 212,812,405 | 0.89 |
| Argentina | 37.62 | 45,851,378 | 0.82 |
| Spain | 37.62 | 49,355,143 | 0.76 |
| Italy | 37.62 | 58,915,656 | 0.64 |
| India | 75.26 | 1,463,865,525 | 0.05 |
Loading the interactive grid…
USD per million people
Showing the top 10 of 24 rows in the current order
From the real demo run: Chinook sales joined to World Bank population (SP.POP.TOTL, 2025). A recorded result, not a live query.
A follow-up question Prototype
What can you say about this data?
Three things stand out. Each one is computed from the rows it points at.
Hover or select an observation to highlight its rows in the grid and the chart. Observations grounded in your dataIn development
Pick the one that fits how you already work. Each card says what to do first and how finished it is.
Claude Code runs the datatug CLI through the skills plugin; other coding agents are planned. The MCP server and MCP web apps (our UIBubbles) are planned.See the skills and how to install them
› /plugin install datatug@sneat-co
✓ Skills for the datatug CLI are installed.
› Which countries buy the most music relative to their population?
Running datatug query run sales/chinook-sales-per-capita
Ireland 8.32 Czech Republic 8.29 Finland 7.37
✓ Done. The result is back in the conversation.
›
Start with/plugin marketplace add sneat-co/ai-marketplace
Ask about your data in plain words, in the terminal or in the web app. It is one session: start in one place and carry on in the other.
Session: chinook-countries
Which countries buy the most music relative to their population?
Joined Chinook sales with World Bank population.
datatug chatDataTug.appThe same session in the terminal and in the web app
Start withdatatug chat
Suggested joins, a plain explanation of a table and a proposed query, inside the query builder and schema discovery. This is planned, not built yet; until then the chat can already propose joins along your schema's foreign keys.
FromInvoice
AI suggested
Join InvoiceLine
Suggested join, by foreign keyInvoiceLine.InvoiceId → Invoice.InvoiceId
About this table: Invoice
One row per sale, linked to a customer.
Until thendatatug chat
A schema says BillingCountry is text. A meaning says it is a country, which other systems know as IE or IRL, and how sure we are of each fact. That is the job of MeaningGraph. Connecting it to DataTug is planned.
Each fact says whether it came from the schema, a person, the data, or a model.
DataTug will read MeaningGraph before it asks a model, and write back what you confirm.
The same country can be a name, an alias, an ISO code. Meaning is what connects them.
Chinook value: "Ireland"
Meaning: Geo.Country
Alias row: "Ireland" → ie
ISO 3166 codes: IE · IRL
World Bank: ie
Select a node to see its evidence.
Chinook value: "Canada"
Meaning: Geo.Country
Alias row: "Canada" → ca
ISO 3166 codes: CA · CAN
World Bank: ca
Select a node to see its evidence.
Chinook value: "USA"
Meaning: Geo.Country
Alias row: "USA" → us
ISO 3166 codes: US · USA
World Bank: us
Select a node to see its evidence.
Chinook value: "Czech Republic"
Meaning: Geo.Country
Alias row: "Czech Republic" → cz
ISO 3166 codes: CZ · CZE
World Bank: cz
Select a node to see its evidence.
The values, the alias table and the population figures are from the real run. The Geo.Country meaning node is a concept: MeaningGraph integration is planned.
Queries, joins and the way you navigate leave evidence. The idea: DataTug turns repeated evidence into knowledge you can confirm, then reuses it without asking a model again.
Invoice.BillingCountry → Geo.Country
Confirmed by you. The next investigation looks this up instead of asking a model.
The design target for a question asked again and again:
Learning from your work is an idea we are testing. The loop is the design target, not shipped behaviour.
DataTug is designed to reach for the lightest dependable mechanism first, and to say which one it used: what is already known, then a small decision model, then a large model, then you.
Schema, confirmed meanings, reference data, saved queries. Free, instant, repeatable.
Jev, a decision model by TypeSafe AI, ranks candidate tables, fields and join paths. The schema then checks the pick.
New questions, plans nobody has seen, explanations in plain language.
You confirm what will be trusted from then on.
Which countries buy the most music relative to their population?
Ranked by Jev, a decision model by TypeSafe AI. The chosen table is then validated against the schema.
Not all knowledge deserves the same trust. DataTug is designed to keep the difference visible on every claim, so you can tell what was declared from what was guessed.
Reused without asking.
Reused, with the evidence in view.
Shown, never trusted silently.
Reuse then needs no model call.
A query is one move in an investigation. DataTug is built around the whole thing: the question, the path you took, the evidence, and what you learned along the way.
A database client
An AI database client
DataTug
They complement each other. Use whichever is quickest for the next step, whether or not you write SQL.
Not only table and column names, but what they stand for.
PlannedJoin a database with an outside dataset in one query.
AvailableWhat you confirmed is there next time, with its evidence.
ExploringWeb, terminal UI and CLI work on the same session.
AvailableChinook says "USA", the population data says "us", and ISO calls the same country USA and "United States". That is a fact to look up, not something for a model to work out again on every question. So keep it as a table: stored once, versioned, reviewed like code.
| alias | country | how it was matched |
|---|---|---|
| Czech Republic | cz | matched by hand |
| Netherlands | nl | matched by hand |
| USA | us | matched by hand |
| Ireland | ie | equals the ISO name |
The three names that needed a person, and one that did not. The mapping is first-class data, not something a model re-infers on every question.
# data/geo/country_aliases/$records/usa.json
{
"alias": "USA",
"country": "us",
"source": "chinook-database Invoice.BillingCountry; manual match: Chinook spells it 'USA'; countries.names.en is 'United States'"
}Countries, currencies, internal ids and standards, status mappings, lookups.
Point-in-time copies for incidents, test failures, migration rehearsals and regression fixtures.
Outside or expensive data kept so it is fetched once, not on every run.
Selected records, suspicious transactions, intermediate joins and temporary mappings.
Persist and version data
Human-readable files in Git: branchable, reviewable in a pull request, reproducible in CI. The demo already reads its country data from an inGitDB directory. Managing project data this way from inside DataTug is planned.
inGitDBAccess and share
One way to reach very different sources. In the demo, OVDB serves the same two databases to the web app. Discovering sources through its Directory is planned.
OVDBThe session belongs to DataTug, not to the interface. Start in the terminal, open the grid in the browser, and come back with the same context. Or the other way round.
$ datatug
Session chinook-countries ● synced
› Which countries buy the most music relative to their population?
✓ Joined Chinook with World Bank population.
o open grid in browser
Session: chinook-countries
Continued from your terminal
Return to the terminal anytime; it is the same session.
One-shot commands for pipelines and scripts, with managed AI built in.
Managed AI in the CLIAvailablePut a live grid or chart on any page.
<datatug-grid> and <datatug-chart> embedsAvailableDataTug AI is not tied to one provider.
Use your own API key with the provider you already trust, in the browser or the CLI.
Browser chat with your own API keyAvailableDataTug-managed AI with no key to set up: in the CLI today, in the web app next. Your prompts, the schema and the query text go to the cloud. Row values do not.
Managed AI in the CLIAvailableManaged AI in the web appIn developmentA Bubble is a small interactive surface, such as a grid, a chart or an inspector, that appears inside the conversation and composes with others into richer tools. UIBubbles is the open standard for them.
Show the top countries as a grid.
Here are the countries, sorted by sales per million people.
| Country | Per million |
|---|---|
| Ireland | 8.3 |
| Czech Republic | 8.3 |
| Finland | 7.4 |
| Canada | 7.3 |
Chart that.
Charted, and linked to the grid.
Why does Ireland map to IRL?
Here is the evidence for that mapping.
This is a hand-built concept of the experience. The Bubble runtime is planned. About UIBubbles
What one stage learns about the data should be there for the next. The aim of DataTug and MeaningGraph is to carry it from first idea to day-to-day support.
Explore the data and domain knowledge that already exist.
Define entities, meanings, identifiers, relationships and mappings.
Developers and agents use what is known, and feed discoveries back.
Incident possibleInvestigate failures, compare environments, keep the edge cases you find.
Incident possibleSupport and agents investigate real behaviour with everything learned so far.
Incident possibleIt can happen while you develop, in test, in staging or in production. An early incident is a gift: it shows a problem before customers do, while it is still cheap to fix.
Each one starts a focused investigation, and what it teaches flows back into the work ahead.
Incidents are high-pressure moments when important truths about systems and data surface. Preserve those truths instead of losing them when the incident ends.
Incidentius integrationPlanned
Helps you investigate while it is happening.
Preserves what was established: meanings, relationships and diagnostic knowledge.
Holds the incident itself: the event, the context and the response.
The idea: discoveries keep their link back to the incident that taught them, and come back in future incidents, in development, in regression tests and in prevention.
Visit Incidentius.comPointing AI at production data should be safe by design. This is exactly what holds today and what is still ahead, claim by claim.
Run DataTug on your own machine, next to your data: queries run locally. With managed AI, your prompts, the schema and the query text go to the cloud. Row values do not.
AvailableScope what a role may read through policy-enforced saved queries, on SQLite and inGitDB sources today. With no policy set, access is unrestricted, and the terminal database viewer does not apply policies yet.
AvailableConnect to a source as read-only, so an investigation cannot change it.
PlannedApprove an investigation step before it runs, instead of trusting a whole plan at once.
PlannedMask sensitive values on the server, driven by what a field means rather than by its name.
PlannedDecide what reaches an external model, down to field level.
PlannedAn audit trail that is the same record as provenance: who or what did this, from which evidence.
PlannedYou met them one at a time above. Each has one job, and none of them is required to start.
Investigate and learn
The workspace where you and AI work through connected data.
The workspaceUnderstand
What fields, values and relationships mean, with the evidence for each.
meaninggraph.ioAccess and share
One way to reach and publish data across very different sources.
openvaultdb.comPersist and version data
Reference data and snapshots as readable files in Git.
ingitdb.comInteract and extend
Grids, charts and inspectors that appear right inside an answer.
uibubbles.orgKeep what incidents teach
Incident context and response, so the learning outlives the incident.
incidentius.comDataTug is early. Here is who it already serves well, and who should wait.
One list, kept in one place. Every status label on this page reads from it, so the page cannot disagree with itself.
Shipped. You can use it today.
Works from start to finish, but not everywhere yet.
Being built now. Do not plan on it yet.
Designed and scheduled, not started.
An idea we are testing. It may change or be dropped.
The demo needs no account and no install. Ask a question, see the answer as a grid and a chart, and open the steps behind it.
Star on GitHubDataTug is open source under Apache-2.0.
