Questions, including the awkward ones.

No. It never writes — not a migration, not a fix, not once. The most it will do is print SQL for you to run yourself.

It also does not rely on its own good behaviour for this: it asks PostgreSQL what the role is permitted to do and shows you the answer, and it refuses to scan through a role the database reports as a superuser.

Your rows never leave. Not a sample, not a redacted sample, not a count of one person’s orders.

When you ask a question in plain language, your question and a list of table and column names go to the model endpoint you configured. The screen shows you the destination, the identifier count, the byte count and the full list of names, and waits for you to agree. If you never ask a question, nothing is sent anywhere.

You need some model endpoint, and it does not have to be OpenAI. Anything that speaks the OpenAI chat-completions shape works, including a model running on your own machine — in which case nothing leaves the machine at all.

You can also use LEDAR with no AI configured. Connecting, scanning, the entity map and the interview all work without it.

The source is available and you can run it now at no cost. You pay your own model provider for the questions you ask, which measured at well under a cent for three questions against a real schema.

A Pro tier — questions answered through our servers, no key to manage — is planned and not yet on sale. The pricing page says exactly where that stands, and prints no price for a thing you cannot buy.

A chat model answers from a schema you paste in. It has no way to run a query, so every number it gives you is a guess dressed as a fact.

LEDAR runs the queries against your database and writes the sentence itself. The model only picks which part of the schema the question is about. Every number you read came from your own data.

Because the code says what was intended and the database says what happened, and the gap between them is the reason you are looking. Reading an entire repository is also slow and expensive for every question you ask.

Unknown, and the product will not claim otherwise.

The question path has been run cleanly end to end three times, against one database and one model. That shows the mechanism functions. It is not a measurement of how often the answer is right, and no such measurement exists yet.

Yes, and the numbers are published. Of 32 prompt-injection attempts fired at the live question path, 8 succeeded.

Aiming at a target outside the menu the product builds is closed. Aiming at the wrong target inside the menu is still open, because a real schema has plenty of legitimate places to point at.

PostgreSQL only. There is no MySQL, SQL Server or MongoDB support, and none is in progress.

Untested. It is an Electron application with no Windows-only logic that we know of, and "that we know of" is the honest size of that claim. Windows is the platform that has actually been packaged and certified.

Not yet. The package passes the Windows App Certification Kit, but the Store identity and the certification notes are unwritten. Today it is a source checkout you build yourself.

Probably not. You already have psql, DBeaver, dbt tests and Great Expectations, and you already know what you are looking for.

This is built for the person who has a connection string, is responsible for the system behind it, and cannot read a schema.

It stays. A scan history is evidence about your system, and removing an application should not destroy it. If you want it gone, delete the ledar folder in your local application data directory yourself.

Yes, including as part of paid work. The Business Source License permits use, modification and redistribution as a dependency; what it forbids is offering LEDAR itself as a competing product or a hosted service.

Each version converts to Apache 2.0 four years after it is published.