Installation

Requirements

Windows, macOS or Linux. No administrator rights: everything installs into your user account.
Python 3.12 or 3.13, the versions Latet is tested on. You do not have to install it yourself: uv fetches one when needed.
At least one supported agent: Claude Code, Codex, OpenCode, Gemini CLI, or a tool that takes an OpenAI-, Anthropic- or Gemini-compatible address (see GitHub Copilot, JetBrains, Zed and other tools).

One-line installer

On macOS and Linux:

curl -LsSf https://raw.githubusercontent.com/Latet-AI/latet/main/scripts/install.sh | sh

On Windows (PowerShell):

irm https://raw.githubusercontent.com/Latet-AI/latet/main/scripts/install.ps1 | iex

The script does two things, in this order:

Installs uv, the Python tool manager, if it is not on your machine yet.
Installs the latest release of Latet from GitHub as an isolated command-line tool, with uv. Local name detection and text in images come with it; their models are downloaded only when you ask. Where they cannot be installed, Latet is installed without them, and the script ends by saying so; a Latet already installed with them is left as it was instead. If an older Latet is running, it stops it, and does not start it again. To upgrade, latet upgrade is the better way. After the script has upgraded a Latet, run latet start, and latet agents apply claude if Claude Code goes through Latet for good, so that it gets the new plugin.

On macOS and Linux, open a new terminal afterwards: the script adds Latet to your PATH for the terminals you open next, not for the one it ran in. If that one cannot find latet either, run uv tool update-shell and open another.

The script changes nothing else: none of your agents’ settings, nothing that starts at login. New to Latet? latet tutorial shows it at work in about 30 minutes. Then run one session through Latet with latet claude or latet codex, as the Quickstart shows, and route your agents for good with latet global-install once it fits your work.

Two environment variables change what the script does:

Variable	Effect
`LATET_NO_MODIFY_PATH=1`	Leave your `PATH` and shell profile alone. If the script installs uv, set `UV_NO_MODIFY_PATH=1` too, for uv’s own installer.
`LATET_SOURCE`	What to install instead of the latest release: a wheel file, or a requirement such as `latet @ <url>`.

Recommended: names and images from day one

Patterns find what has a shape: a key, an email address, an IBAN. Two things have none, and need a model that runs on your computer:

Names and street addresses in free text, such as Emily Carter lives at 42 Mill Lane: the local name model finds them.
The text in screenshots, such as an IBAN on an invoice: text recognition reads it, and Latet covers each value on the image with its fake. Without it, images leave as they are.

The installer installs what both run on. Download their models once:

latet ner install   # the name model, about 1.6 GB
latet ocr install   # the text recognition models, about 31 MB

Or let latet global-install --ner --ocr download them with the rest of its setup. Both run on this computer only: the models come once from huggingface.co, at a pinned revision and checked against pinned SHA-256s, and nothing is sent anywhere. Name detection and Text in images below say more about each.

Upgrade

latet upgrade

It installs the latest release from GitHub as the installer does, local name detection and text in images included. It stops the daemon first (Windows cannot replace files that are in use), installs the new Claude Code plugin if you use Claude Code, and starts the daemon again. A Latet installed without name detection or text in images gets them this way. Where the new release cannot have them, a Latet that has them is left as it was, and one without them is upgraded without them.

Install with an AI agent

Ask your AI agent to follow llms-install.md in the repository. It installs Latet, which changes nothing on its own, then tells you how you can go on: see it at work first with latet tutorial; try it one session at a time, with no setting of your agents changed; have it on for every session; set up your other tools; download the local models. It waits for your choice, and runs latet global-install with your answers only if you choose every session, for example:

latet global-install --yes --profile balanced --agents claude,codex --auto-mode-review block --ner --ocr

Run it in Docker

To keep Python off your computer, run the daemon as a container instead and point your agents at it: see Docker.

Name detection

Names and street addresses in free text, such as Emily Carter lives at 42 Mill Lane, need a statistical model, which Latet runs on your machine. Secrets and structured data such as emails and IBANs do not need it; names do, so it is worth having from day one.

latet ner install   # downloads the model once (about 1.6 GB)

It runs on the ner extra (ONNX Runtime and a tokenizer), which the installer installs with Latet.

The model is OpenAI’s Privacy Filter, Apache-2.0 licensed, in its int8 ONNX build: it reads many languages, Italian, English, French, German and Spanish included. It comes from huggingface.co at a pinned revision, and every file is checked against a pinned SHA-256 before use. While the daemon runs it takes up to about 2 GB of memory.
After the download it runs on your CPU. Nothing is sent anywhere, and Latet connects to huggingface.co only when you run latet ner install.
A running daemon starts using it at once. latet ner status shows whether it is in use; latet ner remove deletes it. To keep it installed but unused, set ner: off in the configuration.

What it looks for depends on the profile; see Detection.

Text in images

Screenshots carry values too: an email address in a terminal, an IBAN on an invoice. Latet can read the text in images on your machine and mask it, covering each value on the image with its fake.

latet ocr install   # downloads the models once (about 31 MB)

The models are PaddleOCR’s PP-OCRv6 small, Apache-2.0 licensed, from huggingface.co at a pinned revision and checked against pinned SHA-256s; they run on your CPU, on the ocr extra (ONNX Runtime and Pillow) the installer installs with Latet. See Screenshots and images.

A local reviewer (optional)

If you run a language model on your own server (llama.cpp, Ollama, LM Studio), Latet can have it read every text before it leaves and mask what the detectors miss: names in passing, health details, a password written in a sentence. Nothing to install in Latet: turn it on in the dashboard’s Rules page, or see Local reviewer.

Run from source

For contributors, or to try changes that are not in a release yet.

git clone https://github.com/Latet-AI/latet.git
cd latet
uv sync
uv run latet version
uv run latet serve      # the daemon in the foreground on http://127.0.0.1:5283; Ctrl+C stops it

Put uv run in front of every command: uv run latet global-install, uv run latet claude, and so on. Two things differ from an installed copy:

Start the daemon yourself. When the daemon is not running, the Claude Code plugin starts it with the latet command, and in a checkout that command is not on your PATH. Run uv run latet start, or let uv run latet global-install start it at login. uv run latet claude starts it too.
Build the dashboard’s web files. In dashboard/, run npm ci and then npm run build (Node.js 22.12 or newer). Until then, latet dashboard opens a page that says the files are missing, and latet tutorial warns that its page will not show. A Latet installed from the repository’s address (git+https://...) lacks them too.

Start at login runs the daemon from the checkout’s virtual environment, so run uv run latet global-uninstall before you move or delete the checkout. See Contributing for the development workflow.