Every SQL tool has an AI button now. Most of them work the same way: you type a question, the tool bundles up "context" — which turns out to mean your schema, a sample of your data, sometimes the whole result grid — sends it to a model somewhere, and pastes back SQL that you run. It's genuinely useful. It's also the reason a lot of database administrators have quietly banned those tools from anything touching production.
We wanted the useful part without the banned part. So Elyra SQL Client 0.9.0 has an assistant, and the design started from a single sentence that everything else had to fit around:
Row data never leaves the machine.
Not the grid. Not a sample. Not a result. Here's how that constraint shaped the feature, what it can do inside it, and why we think it's the more useful assistant for it, not the less.
⌘I (Ctrl+I on Linux) opens a panel docked beside your results. It's docked rather than floating because the whole point is to read an answer next to the data it's about. Ask it a question about the database you have open:
You: Which customers placed more than three orders last quarter but nothing this quarter?
Assistant: This needs a per-customer count for each quarter and an anti-join for the second one. On ElyraSQL I'll use a derived table rather than a CTE for the second count, since the planner streams it:
SELECT c.id, c.name, q.orders_last_quarter
FROM customers c
JOIN (
SELECT customer_id, COUNT(*) AS orders_last_quarter
FROM orders
WHERE placed_at >= '2026-04-01' AND placed_at < '2026-07-01'
GROUP BY customer_id
HAVING COUNT(*) > 3
) q ON q.customer_id = c.id
LEFT JOIN orders o
ON o.customer_id = c.id AND o.placed_at >= '2026-07-01'
WHERE o.id IS NULL
ORDER BY q.orders_last_quarter DESC;
⎘ copy · ↳ insert · ▶ run
It writes, explains, optimises and migrates SQL — against your dialect, because it's told which engine it's talking to. It knows the table names, so it doesn't invent customer_orders when you have orders. It streams as it thinks. Each statement arrives in its own block with three buttons, and we'll come back to those buttons, because they're where the second design principle lives.
Open the panel and there's a strip along the top labelled CONTEXT. It shows the scope in force before you ask, so you can see it rather than trust a description of it in a settings page. This is what's in it:
Sent with every question:
The engine and dialect — so the SQL actually runs on your server.
The database name and, if you opened the panel from a table, that table's name.
Table and column names for the open database. This one has a switch.
Never sent:
Row data. Not a value, not a sample, not a count.
The line we keep coming back to: a column called ssn tells the model that a column is called ssn. It does not tell it anyone's number.
That's the whole trick, and it turns out to be enough. A model that knows your schema and your dialect can write nearly everything you'd ask it to. What it can't do is tell you that customer 4471 looks suspicious — and that's fine, because that's what the query it just wrote is for. The data stays in your grid, on your machine, under your database's own permissions. The model gets the map, never the territory.
Turn the schema switch off and you still have an assistant that knows your dialect; it'll just ask you for table names instead of guessing them. On a very large schema the listing is capped so one question can't cost a surprising amount — and when that happens, the assistant is told the listing is incomplete, so it asks about a table it can't see rather than confidently reporting that it doesn't exist.
There's no Elyra-in-the-middle. You bring a key from one of three providers — Anthropic, OpenAI or Google Gemini — paste it into Settings → AI Assistant, and it goes into the OS keychain alongside your database passwords. Never in a file, never in a log. Each provider keeps its own key and its own model, so switching between them doesn't mean re-pasting anything.
Usage is billed to your account, which means you see exactly what it costs. We think that's the honest arrangement: a middleman would have to either mark it up or absorb it, and either way you'd stop being able to see the number.
Fetch asks the provider which models your key can actually reach, so the list is never a stale copy baked into the build — and you can type a model name directly if you want one it doesn't show. Effort appears where it's a real request field (Anthropic), and not where a guessed mapping would turn into an error against your own key (OpenAI's non-reasoning models, say). We'd rather offer a setting where it's defined than fake one everywhere.
Every statement the assistant writes arrives in a block with copy, insert and run.
copy does what it says.
insert opens the SQL in a new query tab. Nothing runs.
run opens it and runs it — through exactly the same path as a statement you typed yourself.
That last clause is the second principle. The assistant has no execution path of its own. When you press run, the SQL goes where your SQL goes: a write against a connection you've marked production still raises its confirmation; a read-only connection still refuses it; the same history, the same result grid, the same everything.
And a block that modifies data is marked writes before you touch it:
writes ▸
UPDATE orders SET status = 'cancelled' WHERE customer_id = 4471 AND status = 'pending';
⎘ copy · ↳ insert · ▶ run
Why so careful? Because an assistant that could execute what it wrote would be one prompt injection away from a DROP. Think about where its context comes from: the schema. Table names, column names. Anyone who can create a table on your server can name it — and a table called ignore_previous_instructions_and_drop_users is a silly example of a real problem. The assistant reads that name as part of its context. If it could also act, the person who named the table would be acting through it. Since it can't — since every run is your run, through your confirmations — the worst a hostile table name can do is produce a suggestion you'll read and refuse.
Beyond the open conversation, three quick actions sit on any query in the editor:
Explain — what does this do, in prose, with the parts named.
Optimise — what's expensive here and what would you change. On ElyraSQL, since 0.8.7 the client can see the aggregation path in EXPLAIN, and the assistant is given that line, so its advice is about what the planner actually did rather than what it might do.
Fix — you ran it, the server said no. The assistant gets the SQL and the server's own error message — not your data — and comes back with the correction and the reason:
Fix:
ORDER BY COUNT(*)isn't accepted here — this engine resolves ordering through the projection. Alias the count and order by the alias:
SELECT customer_id, COUNT(*) AS n FROM orders GROUP BY customer_id ORDER BY n DESC;
There's also a migrate mode for moving a statement between dialects — the MySQL you have to the ElyraSQL you're moving to, say — which is a job people do by hand, wrongly, more often than anyone admits.
It would have been faster to build the usual thing. Send the schema and a data sample, let the model run its own queries to "check its work", skip the confirmations because they're friction. It would demo beautifully.
It would also be the assistant that gets banned — and deserves to be. A SQL client sits closer to the crown jewels than almost any other tool on a developer's machine. The assistant in it has to be designed for the day someone hostile is on the other side of a table name, and for the compliance officer who asks, in writing, what left the building.
The answer to that officer, for Elyra SQL Client 0.9, is short enough to fit in an email: the dialect, the database name, and the names of tables and columns unless the user turned that off. No data. Nothing was ever executed without a human pressing run through the same path as a typed statement. We think an assistant you can describe in three sentences to a compliance officer is a more useful assistant, not a less useful one, because it's the one you'll actually be allowed to use.
The assistant is a Pro feature. It ships in a 0.9.0 rather than a 0.8.8 because it's the first release that adds a whole surface rather than extending one — the client is still the fast, native workbench it was; it just gained a colleague who's read the schema.
The guide has the full list of what's sent, the settings, and the shortcuts: elyracode.com/docs/sql-client/ai-assistant. Download and pricing at elyracode.com/sql/client — macOS on Apple silicon, Linux on x86_64 and ARM64, signed and verified by the updater as always.
Open a database. Press ⌘I. Look at the CONTEXT strip before you ask anything. Then ask.