Questions, answered

For a data team that already has a desktop BI licence, it replaces the "export a CSV, paste it into ChatGPT, screenshot the chart, rebuild the deck" loop: one Windows app holds the warehouse, writes the SQL, draws the chart and exports a real Power BI .pbix file. It does not replace a hosted multi-user BI service — there is no shared server, no live URL and no commenting on a dashboard. If you need ten analysts querying one cloud warehouse simultaneously, keep your cloud BI and use DataAgent for the offline, on-premise and .pbix work.

A Windows 10 or 11 machine, 64-bit. One download: DataAgent-Setup.exe (roughly 439 MB). The Python runtime, the DuckDB engine and the interface are inside the installer, so there is nothing else to install first and no developer tooling involved. Budget roughly 0.7–1 GB of disk once installed, plus your own data, and about 4.7 GB more if you add a local model. The download page has the full system requirements.

The packaged download is Windows-only today. The project is open source under the GNU AGPL v3 licence, so the same stack runs from source on macOS or Linux if you are comfortable with a Python environment — but there is no packaged build, and the desktop shell and installer are Windows-specific.

No account anywhere but inside the app: you register the first user yourself when the app opens. No card is needed to use the free tier. You only need an API key if you choose a hosted model provider; a local model through Ollama needs no key and no account at all.

This release is not code-signed yet, so SmartScreen shows “Windows protected your PC” with an “Unknown publisher” warning. Choose More info, then Run anyway. The portable build (DataAgent.exe) sidesteps the installer entirely — it runs from any folder, including a USB drive.

There is a free tier and there are paid tiers — one price list, in rupees: Free is ₹0, Pro is ₹499/mo, and Lifetime is ₹10,000 once. The free tier gives you 25 AI requests a day, 3 saved dashboards and 7 days of chat history, with chat, SQL, file upload and the spreadsheet workbench available. The paid tiers raise the daily allowance and unlock the export, automation, live-connection, governance and enterprise tools. The pricing page has the full comparison. The source is open under AGPL-3.0, so you can also read it, fork it and run it yourself.

No. A local model runs on the CPU, just more slowly. A GPU is optional and only changes how fast the model answers. If you would rather not run a model at all, SQL, the spreadsheet workbench, dashboards and exports work without one — the assistant then answers from rules and tells you it did. Every hosted provider works on any machine, but your prompts go to that provider.

Checkout runs on Razorpay (UPI, cards, netbanking) or on the separate billing site, depending on the deployment. The app never sees your card details, and the plan is attached to the email you pay with — there is no code to paste. For a refund, write to us with the email you paid with and the payment id; the billing provider processes the refund and the plan is revoked automatically when it clears. We have not published a blanket refund window because a paid licence is a one-off purchase, and we would rather look at your case than argue about a rule.

A small team, no SLA. Bug reports in the contact page and on the issue tracker get answered; feature requests get read. Paid plans are described as "priority support" in the plan table, which means they are looked at first — not that a response is guaranteed within a fixed time. Because the source is open, most questions are faster to answer by reading the code.

Yes. DataAgent is open source under the GNU AGPL v3 licence: use it for personal projects, internal tools and client deliverables. The AGPL condition is that if you modify it and let other users interact with it over a network, you must offer those users the modified source. Running it unmodified for your own team, including self-hosting it, triggers none of that. The full terms are in the LICENSE file at the root of the repository.

The desktop app ships no telemetry, no analytics, no crash reporting and no background sync. The warehouse is a local DuckDB file and the app runs on your machine. Outbound traffic goes to three places, and only these: the model provider you configure (choose Ollama or LM Studio for a fully offline setup); GitHub, for the update check; and, on paid plans only, the entitlement service and the proxy pool behind the IP-rotation feature. The security page lists every connection, what triggers it and what data it carries.

In %LOCALAPPDATA%\DataAgent — the warehouse, your uploads, the logs and your settings, in one folder you can back up by copying it. The application itself lives in Program Files. Your data folder survives an uninstall.

No. Analysis opens the warehouse read-only and all SQL is parameterized. Writes happen only through explicit ETL runs handled by a single dedicated process, because DuckDB allows only one writer per file.

Nobody, unless you configure a hosted model provider — then that provider sees your question and the schema context it needs to answer it. If your data cannot leave the building, use a local model. Single sign-on and row-level security are implemented but ship disabled by default, so if you rely on them, switch them on deliberately rather than assuming they are running.

Files upload directly (CSV, TSV, TXT, JSON, JSONL, NDJSON, Parquet, XLSX and XLS) and are ingested into the local DuckDB warehouse. Live database connections are SQLite, PostgreSQL and MySQL, configured in db_connections.json. There are no cloud object-store connectors: no S3, no GCS, no Azure Blob. Dashboards can be imported from and exported to Power BI (.pbix).

They are honest about their scope. Regression Explorer compares a linear fit against a polynomial one and reports adjusted R²; it is not a model search, and there is no cross-validation or ensembling. Segment Breakdown groups rows and aggregates them — it is a group-by, not clustering. Trend Forecast projects a least-squares line and decomposes a series — it is not ARIMA, ETS or Prophet. The names describe what the code actually does.

Not in this build. The PDF export produces a plain text summary of the table, because the HTML-to-PDF renderer is not bundled into the installer. The fully formatted exports are Excel, PowerPoint, HTML, Markdown, CSV, Parquet and the Power BI .pbix — use those when the layout matters.

Yes. The runtime, the interface, the warehouse and ETL all ship inside the installer. For AI chat you additionally need a local model: the app can install Ollama for you in one click from Settings, and it then downloads a model once (roughly 4.7 GB for the recommended qwen2.5:7b). After that, the network is only needed to pull a model or to call a hosted model provider.

Yes — the same FastAPI backend that serves the interface, on 127.0.0.1:8002, with REST endpoints for auth, chat, SQL, data, dashboards and export. The API reference in the docs lists the real paths.

Because it is embedded: no server to run, no cluster to pay for, and the whole warehouse is one file you control. It is an analytical engine with columnar speed, which is the trade-off local-first requires. The cost of that choice is scale: one laptop DuckDB will not match a warehouse cluster on a five-billion-row join.