Spreadsheet fan?
Edit a live grid. Join datasets, create pivot tables, generate new columns. Each edit becomes a step in the pipeline , so it's recorded, not just applied. Apply your spreadsheet skills to large datasets.
Analyze anything. Approve every step.
A spreadsheet, a notebook and a dashboard over one dataset and one history. The agent plans, and you edit or approve it before anything executed. Bring your favorite model - your own cloud key, something small running locally, or no AI at all.
Free beta · Fully usable with the AI switched off
Notebook
A notebook over the same history. Every step is a cell you can read and re-run — the record of how the dataset got here.
Dashboard
Share your results easily,
One dataset, one history. The dataset's state is a replay of an ordered list of steps — which is why undo, redo, save, load and audit are all the same mechanism rather than four features.
A plan, not code
The agent hands you a plan, dry-run and drawn as a flowchart with per-step results. Edit it or approve it — nothing commits until you do.
Runs again next month
Save an analysis and replay it on a new file. The step list is the pipeline, so re-running it isn't a rebuild — it's the same steps again.
Execution stays on your machine
Prompts carry your schema, profile statistics and the plan — never your rows, in any AI mode. There's no data server to send them to.
| Mode | What it means |
|---|---|
| No AI | Turn the agent off. Everything else still works — it's a data IDE on its own, not a chatbot with a grid attached. |
| Local | Run a small model on your machine. Free, and enough for routine steps. |
| Your key | Bring your own cloud API key. You hold the billing relationship with the provider — we're not a reseller. |
| Frontier | Claude and other frontier models, for the steps where the reasoning is genuinely hard. |
Switch whenever you like. Because a prompt carries your schema and profile statistics rather than your rows, the cost of a step tracks how wide your data is — not how many rows are in it.
Full access. Run it with no AI, with a local model, or with your own cloud key — any model spend is yours, billed by your provider.
Paid plans arrive once the beta wraps — shaped by what beta users actually need.
DataGadt is for working analysts and data scientists doing the large share of analysis that fits comfortably on one machine — not a cluster, not a procurement cycle. Good work there kept getting stuck between three tools, and AI-assisted steps meant handing over control: generated code you didn't ask for, run against your data, with no record of how you arrived anywhere.
So the agent here proposes a plan you read and approve, the three views share one dataset and one history, and the model behind it is a setting you own — including off. What you get isn't a faster black box; it's an analysis you can hand to someone else and have them see exactly how it was built.
Download the beta
Beta access is by invite — log in if you have an account, or request one below.
The beta isn't code-signed yet, so Windows may show a "protected your PC" notice — click More info → Run anyway.