Data · the long version

Connect a database.
Ask in plain language.

Wire up a real database, then ask about it like you'd ask a colleague. The model can see your schema, so it writes read-only SQL, runs it, and charts and explains the answer — or hands you the query to run yourself. Never paste a CSV again.

Data sources

A SQL workspace for real data.

Connect a database and you get a real workspace: its tables in a sidebar, SQL in tabbed editors, one click to run, and a results grid you can export. No exports, no stale snapshots — the live database, queried at chat speed.

connect anything

Six engines, plus files.

PostgreSQL, MySQL, SQL Server, Oracle, SQLite, and DuckDB — connect the database you already run. Or drag in a SQLite or DuckDB file and query it on the spot, with no server to connect to.

schema-aware AI

Draft a query, or let it run.

Because the model can see your tables and columns, you can skip writing SQL. In Draft, it writes a query and drops it in your editor to review before you run it. In Agent, it runs read-only queries itself, works through the results, and explains what it found.

read-only by design

It can read. It can't write.

Only SELECT runs. Anything that would change data or structure is refused before it reaches your database, and every result is size-capped. Credentials are stored encrypted, and you can connect over your private network — internal databases stay internal, with no public endpoints.

reuse everywhere

Export it, or run it again.

Send any result out as CSV or Excel, or copy it as TSV straight into a sheet. And a query you write is reusable: pull it into a chat to keep digging, or drop it into a workflow as a step that runs on every run.

catalyst · data sources
finance_db · postgres prod · 14ms · healthy
orders · mysql prod · 22ms · healthy
crm · sql server replica · 9ms · healthy
warehouse · oracle private net · 140ms · slow
ledger.duckdb · file uploaded · 2ms · healthy
SELECT region, SUM(amount) AS revenue FROM orders WHERE quarter = 'Q2' GROUP BY region ORDER BY revenue DESC
Knowledge attachments

Drop a PDF on a prompt.
The model just knows it.

Some context isn't in a database — it's in a PDF, a deck, a spreadsheet. Attach files to a Prompt and they become part of what it knows, in every session, with no copy-and-paste.

how it reads

Small files inline, big files on demand.

Small files are read in full, every turn. Larger documents are searched by meaning and pulled in only where they're relevant — so a 200-page report doesn't crowd out everything else.

file types

PDF, Word, PowerPoint, Excel, CSV, JSON.

Upload the formats your work actually arrives in. Each is turned into clean, readable text the moment it lands — and the original file stays available for any analysis that needs it.

stays focused

Only this prompt sees it.

Documents you attach to one Prompt belong to that Prompt. Other prompts and sessions don't see them, so the model's attention stays on the task in front of it.

real-world limits

15 MB per file · 50 files per prompt.

Sized for real documents, not toy ones — annual reports, technical specs, and customer transcripts all fit comfortably.

safe sharing

Share the prompt; keep the files.

Share a prompt and people see what it knows, but they can't download the underlying files. Your storage is never exposed by a share link.

clean delete

Delete the prompt, delete everything.

Removing a Prompt removes its files and everything derived from them in one step. Nothing is left behind.

Your data, your model

Stop copy-pasting.
Start asking.

Real connections to real databases, read-only by design. Real documents attached to real prompts. The model sees exactly what it needs, when it needs it — and nothing else.