Understanding a legacy codebase with AI, without exposing the code
An inherited repository with no documentation, an urgent bug, nobody left to explain it. Adding the project folder to a space saves days, provided you check seven points first.
Monday, first day on the job. The previous contractor has gone, leaving behind a repository of fourteen hundred files, a three-line README and a ticket marked urgent: "since the update, some invoices come out with the wrong VAT rate". Nobody on the team knows where that rate is calculated.
AI saves days on this kind of work: following a call from file to file, explaining a function with no comments. But you still need to look at what the folder contains before you hand it over.
Legacy code: code that isn't yours
Inherited code usually belongs to the client or the employer. It falls under trade secrets, which the law protects in Switzerland as across Europe, and your contract generally includes a confidentiality clause. A legacy project also hides files nobody ever sorted, sometimes copies of the production database, whose personal data falls under the GDPR or the Swiss Federal Act on Data Protection (FADP).
Try it on your own caseYour notes stay in your browser, never on our servers.
Show me the service
No account needed
First month free
The checklist before adding the project folder
Tick each point once you have removed it, or checked it, in what you are about to paste. The sorting is yours: nothing is automatic, nothing is stored.
Everything is sorted. Paste the rest: the AI needs nothing else. Try it on your own case
Understanding the legacy project, one question at a time
With a signed-in account, the assistant automatically searches the added folder for the passages relevant to each question; the index stays in the folder, on your computer. How much code can be indexed depends on your plan: see the plans. Start with an overview:
Using the files in the folder, map out the project: the main modules, the entry points (routes, scheduled commands, queued jobs), the dependencies between them and the code that looks dead. Finish with the five questions I should ask the previous team.
Then follow the problem you are working on, and ask for evidence:
Find where an invoice's VAT rate is calculated, from the "Download invoice" button all the way to the PDF. Give the file paths in call order, quote the lines that set the rate, and separate what you read in the code from what you are assuming.
That last instruction matters: AI gets things wrong with confidence, invents a method or describes the code it expected rather than the code that exists. Open every file it cites, check every line, and run any fix through your tests and the team's review. For reviewing code, see the mistakes that leak proprietary code.
What stays confidential
With a confidential model, the relevant passages of code leave your machine to be processed on servers in Switzerland and Europe, and then nothing is kept on our side. A folder marked "Confidential models" is never sent to an anonymised model, and a conversation that used it stays on those models.
A folder without this option is handled the way you chose: with an anonymised model, its content goes out as it is. The Confidentiality filter, for its part, examines your messages and lets you decide before sending; it masks names and identifiers, not the logic of an algorithm, as the guide on anonymisation and pseudonymisation explains.
The history of your conversations lives in your browser. On a shared workstation or a common staging machine, anyone using the same session can read it: sign out when you leave. And a secret removed from the folder is only protected if it is also removed from the repository's history; the list of data never to paste into AI applies to a folder too.
To get a first module explained, secrets removed: open the chat.
