Prove read-only, do not promise it.
LEDAR asks PostgreSQL what the role is permitted to do and shows the answer. A role that turns out to be a superuser is refused, and there is no button to override that.
An AI employee on your own machine
Know what is inside.
You inherited a database you did not build. LEDAR reads it without writing to it, explains what is in there in plain language, and tells you exactly where its knowledge stops.
L
LookRead the real system, not a description of it.
E
ExplainSay what is there in a sentence a non-engineer can act on.
D
DiscloseName what it did not look at, every time it reports.
A
AdmitName what it looked at and still does not understand.
R
RetainKeep the history, so knowledge does not leave with a person.
LookExplainDiscloseAdmitRetain
The situation
You are responsible for a system nobody explained to you.
Not a data engineer. Not looking for a query tool. You have a connection string and a system that has to keep working.
What every other tool assumes
- You already know which table to look at.
- You can read a schema and spot what is missing from it.
- You know the question worth asking before you open the tool.
- Someone is around to tell you what the columns mean.
What is actually true
- The developer or the agency is gone.
- The product was assembled with AI and works, and nobody mapped it.
- A layoff left one person holding a system they did not write.
- You find out a rule was broken when a customer tells you.
Pointing a language model at the whole codebase is the obvious idea, and it is the wrong one: it is slow, it costs a lot, and the answer it gives you comes from code rather than from the rows that actually exist. LEDAR reads the running database instead, then answers from what it measured there.
How it works
From "I inherited this" to a bounded answer.
Five fixed questions, one scan, and a map. After that you ask in your own words and get a timeline back.
Scan, and build a map.
Constraints, relationships, row counts and sizes. Foreign keys the database declares, plus implicit links it measures and marks as measured — never a link it merely guessed from a column name.
Ask in your own words.
Type "I paid but no order appeared". LEDAR walks the map from the customer to the payment to the order and answers with a timeline, naming the point where the trail breaks.
The product contract
Trust needs edges.
DISCLOSE
Every answer states what was not looked at. A negative claim with no coverage boundary is refused rather than softened.
ADMIT
Every answer states what it looked at and still cannot explain. A route it could not afford to walk is named, not dropped.
PROVENANCE
What the database enforces, what you told it, and what it merely observed never look alike on screen.
READ-ONLY
It never writes to your database. Not a migration, not a fix, not once. The most it will do is print SQL for you to run yourself.
What leaves your machine
Your rows never leave.
Not a sample. Not a redacted one. Not a count of one person’s orders.
When you ask a question, LEDAR sends your question and a list of table and column names to the model you configured, so it can pick which part of the database the question is about. That is schema, not data — but it is not nothing, and calling it nothing would break the two letters in the middle of the name.
- The screen names the destination, the number of identifiers, the byte count and the full list of table names — and then waits for you.
- The permit is hashed over the exact bytes to be sent. Change one identifier afterwards and the send fails instead of proceeding.
- If you never ask a question, nothing is sent anywhere. Scanning, the map and the interview all work with no AI configured at all.
- You supply your own model key. It is stored with the operating system’s own encryption, and if the OS cannot encrypt it, LEDAR refuses to store it rather than writing it to a file in the clear.
Measured, not asserted
Numbers with a tool behind them.
Hover any figure on this site and it will name the command that produced it and the day it was run.
- Tests, 0 failing, 0 skipped
- 1,182
- Tables read on the Pagila sample
- 94
- Rows on the largest database it has run against
- 344.3M
- Windows certification checks passed
- 23
The scanner has been run against a 344 million row MusicBrainz copy and a 12 GB crates.io copy, not only against a toy sample. On MusicBrainz it found that all 758 declared foreign keys are marked NOT VALID — meaning the database is not enforcing a single one of them. That is the kind of thing nobody types a query to look for.
Admit
What is not done.
A product whose middle letters are Disclose and Admit does not get a landing page that only lists wins.
- The question-answering layer has been run cleanly end to end 3 times, against one database and one model. That shows the path works. It is not a measurement of how often the answer is right.
- We fired 32 prompt-injection attempts at that same path and 8 of them got through. The published defence closes attempts to aim outside the menu; aiming at the wrong table inside the menu is still open.
- Not in the Microsoft Store yet. The package passes certification, but the Store identity and the certification notes are unwritten, so today this is a source checkout rather than an installer.
- PostgreSQL only. macOS and Linux are untested — the code has no Windows-only logic that we know of, and "that we know of" is the honest size of that claim.
The full list, including the measurements that came out badly, is on the status page.
Run it
Three commands, then a connection string.
You need Node 22 or newer and a PostgreSQL connection string. There is no account and no sign-in.
git clone https://github.com/ledar1106/ledar-src.git
cd ledar-src
npm ci
npm run desktopThe first launch downloads the Electron binary, about ~100 MB, once. Then you paste a connection string and LEDAR proves the connection is read-only before it reads anything.
Find out what is actually in there.
Source-available, runs on your machine, and tells you what it did not look at.
Get the source
Or read how it works first.