Tutorial

latet tutorial shows Latet at work in your browser, in about 30 minutes, before you point a real agent at it. A pretend agent sends its requests to a pretend model. Latet, between them, is the real one: the same masking, checks and dashboard you get with your own agent. The data is invented, and nothing leaves your computer.

Start it

latet tutorial

It opens the tutorial in your browser, on the menu of chapters, and runs until you end it. The terminal says where it is:

Tutorial at http://127.0.0.1:52817 — Ctrl+C ends it

Command	What it does
latet tutorial	Opens the menu of chapters.
latet tutorial terms	Opens one chapter, by its id: here, Your own terms.
latet tutorial --list	Prints the chapters, their minutes and which you finished, then stops.
latet tutorial --print-url	Prints the sign-in link instead of opening the browser. It works once, for two minutes.
latet tutorial --keep	Keeps the sandbox’s folder when the tutorial ends, and says where it is.

The tutorial runs on your computer, not in Latet’s container.

The chapters

Each chapter takes a few steps. You send the agent a prompt, typed for you, and Latet’s pane shows what happened to its request, in the dashboard’s own views. Some steps ask you to act there: click a value, add a word to your terms, approve a request, save your guidance.

Chapter	Minutes	What happens
basics: The basics	5	A key in .env, an email address and a phone number leave as fakes; the file the agent writes holds the real key.
terms: Your own terms	3	A codename leaves in clear until you add it to your terms; from then on the model gets a fake, a name it has never seen.
false-positives: A value masked for nothing	3	Stripe’s test card is masked like a customer’s card; Not sensitive lets it through.
stopped: A request stopped	3	A key pasted in a conversation from before Latet was on stops the request; a new conversation without it goes through.
live: The Live Reviewer	3	A request waits for your approval: you mask a customer’s name by hand, then approve it.
settings: An agent lowering your rules	3	The agent asks Latet to let email addresses through, and the change waits for you: approve it or undo it.
data: Production data	4	Strict data mode masks every value of a query’s table; you keep ids, dates and plans in clear.
llm: The local reviewer	4	A support ticket holds an email address written out, an IBAN with a typing mistake, a name and a diagnosis: no pattern finds them, the local reviewer does. Then your guidance has it find the hospital’s wards too. The reviewer is simulated: the Tutorial Reviewer.
images: Screenshots	3	The agent sends a screenshot of a refused payment: Latet reads the email address, the IBAN and the phone number in it, and covers each with its fake before the image leaves. The chapter needs text recognition: latet ocr install.

A chapter ends with what you learned and where it is in the dashboard. The last one ends with how to use Latet with your own agent. The menu marks the chapters you finished, from one run to the next, and Start over puts a chapter back as it began.

A sandbox

The chapters mask values, stop requests and change settings, so they run on a Latet of their own, in a sandbox:

• Its own data, settings and history, in a temporary folder, and its own server on 127.0.0.1, on a free port. Your own daemon is not touched, and need not be running.
• Nothing of yours changes: your settings, your vault and history, and your dashboard’s sign-in (the tutorial’s page signs in with a cookie of its own).
• Nothing leaves your computer. The sandbox’s one route leads to the pretend model, which runs inside the latet tutorial command: no request can reach a provider, and no API key or sign-in is needed.
• The same fakes at every run: the sandbox makes them from a fixed key, not from yours.
• One file outside it: tutorial.json in your data folder, with the ids of the chapters you finished.

The sandbox has no name detection and none of your identity: the chapters show what the detectors find, and what you mask by hand. In the local reviewer’s chapter, the reviewer is the Tutorial Reviewer: simulated, it finds what the chapter says, and no request reaches a server. In the screenshots chapter, the sandbox reads the text in images with your own text recognition models, once they are installed (latet ocr install). Every other page of its dashboard is there to explore, with a bar on top that brings you back to the chapter.

End it

Click End the tutorial in the page, or press Ctrl+C in the terminal: the sandbox stops, and its folder is deleted. A run cut off another way, its terminal closed for example, leaves the folder in your temporary folder, and the next latet tutorial deletes it. A run that stops on an error prints the last lines of its log, and keeps the folder, with the whole log, until the next run.

In your own dashboard

What the chapters show, in the dashboard of your own Latet (latet dashboard):

What	Where
Every request: your computer’s text, and what the model got	Live, then the request
Every value Latet replaced, and its fake	Vault
Your terms	Rules → Terms
The values you let through	Rules → Allowed values
The requests Latet stopped, and why	Live, filtered by Blocked, then the request
What waits for your approval: requests, and changes to your settings	Approvals
The Live Reviewer, for every project or for one	Rules → Reviewers → Live Reviewer
Strict data mode, and the columns you keep in clear	Rules → Protection → Production data
What the local reviewer found in a request	Live, then the request: Local reviewer
The local reviewer, and your guidance	Rules → Reviewers → Local reviewer
An image a request carried, as it was sent, and the text read in it	Live, then the request: Images
What leaves with an image or a PDF, and whether the text in images is read	Rules → Protection → Production data: Images and PDFs

Then try it on your own agent, one session, with a made-up key: see the Quickstart.