Docker

Latet also runs as a container: no Python on your computer, and the daemon kept apart from the rest of it. The image holds the daemon, the dashboard and the runtime of the optional name model and text recognition; their models are downloaded only if you ask. Your agents stay on your computer and talk to the container on 127.0.0.1.

The container cannot change your agents’ settings or see who you are, so two steps that latet global-install does by itself are done by hand here: pointing the agents at it, and telling it your identity. Both are a copy and paste.

Start it

Save compose.yaml from the repository, then:

docker compose up -d

Without Compose, the same in one command:

docker run -d --name latet --restart unless-stopped -p 127.0.0.1:5283:5283 \
  --read-only --tmpfs /tmp --cap-drop ALL --security-opt no-new-privileges:true --init \
  -v latet-data:/data -v latet-config:/config ghcr.io/latet-ai/latet:latest

docker exec latet latet status says whether it is up. The container runs as a user without privileges, on a read-only file system, and keeps everything in two volumes: latet-data (the vault, its key and the dashboard history) and latet-config (config.yaml).

Point your agents at it

docker exec latet latet agents print

It prints, for each agent, what to change and a command that does it. On Windows, add --shell powershell. If an agent’s settings file does not exist yet, create it empty first, such as mkdir -p ~/.codex && touch ~/.codex/config.toml. In short:

• Claude Code needs the plugin and a few variables in ~/.claude/settings.json:

  claude plugin marketplace add Latet-AI/latet
  claude plugin install latet@latet
  docker exec -i latet latet agents merge claude < ~/.claude/settings.json > ~/.claude/settings.json.new \
    && mv ~/.claude/settings.json.new ~/.claude/settings.json

  agents merge reads the file, adds the variables and writes it back, so the token the plugin shows the daemon goes into the file without being printed. One of the variables, LATET_AUTOSTART=0, tells the plugin not to start a daemon of its own: if the container is down, Claude Code says so and withholds requests until it is back.

• Codex needs openai_base_url in ~/.codex/config.toml, which this writes (--codex-login api-key if Codex signs in with an API key instead of ChatGPT):

  docker exec -i latet latet agents merge codex < ~/.codex/config.toml > ~/.codex/config.toml.new \
    && mv ~/.codex/config.toml.new ~/.codex/config.toml

• OpenCode needs a baseURL for each provider in opencode.json, the same way.

PowerShell has no < redirection, so agents print --shell powershell copies the file into the container’s latet-data volume with docker cp, merges it there, and copies it back.

Restart the agents afterwards. For these routes the modes and the protection are the same as with the daemon installed on your computer: Anthropic for Claude Code, OpenAI or ChatGPT for Codex, and OpenCode’s providers. Claude Code on Bedrock, Vertex AI or Microsoft Foundry, a Codex provider of its own and a company’s gateway are set up by latet claude, latet codex and latet global-install, which run on your computer: use Latet installed on your computer for them.

Tell it who you are

Latet masks your user name in paths, your host name and your git author. In the container it cannot see them, so tell it once, from your computer (this is also in agents print):

docker exec latet latet identity set --user-name "$(id -un)" --user-name "$(basename "$HOME")" \
  --host-name "$(hostname)" --full-name "$(git config --global user.name)" \
  --email "$(git config --global user.email)"

docker exec latet latet identity show prints what it masks.

Everyday use

What	Command
Open the dashboard	docker exec latet latet dashboard --print-url, then open the link on your computer
Check everything	docker exec latet latet doctor
Download the name model (about 1.6 GB, once)	docker exec latet latet ner install --yes
Download the text recognition models (about 31 MB, once)	docker exec latet latet ocr install --yes
Your own recognizers	docker exec -i latet latet rules add - < rules.yaml
Update	docker compose pull && docker compose up -d
Another port	LATET_PORT=6655 docker compose up -d, then agents merge again
A backup of the settings and the vault	docker exec -it latet latet backup /data/latet.backup --vault, then docker cp latet:/data/latet.backup . (restoring it)

The settings live in config.yaml in the latet-config volume; the dashboard’s Rules page changes them, and docker exec latet latet reload applies a file you edited.

Projects and the local reviewer

• Projects: your own settings for a project work by its folder’s path. A project’s latet.yaml is read only when the folder is mounted in the container at the same path.
• The local reviewer runs on your computer, outside the container, and in the container 127.0.0.1 is the container itself. Set its address to http://host.docker.internal:8080/v1, which Latet accepts in its container only. On Linux, add extra_hosts: ["host.docker.internal:host-gateway"] to the service in compose.yaml, and have the model server listen on an address the container can reach.

Security

• Publish the port on 127.0.0.1 only. On Linux, Docker’s published ports go around the firewall: -p 5283:5283 would open Latet, and the requests your agents send through it, to your network. compose.yaml and the command above publish on 127.0.0.1.
• Other containers on the same Docker network can reach the port. The bridge, the dashboard and the admin routes still need the token or a dashboard session, and every request must name 127.0.0.1 or localhost as its host; the proxy routes forward the caller’s own credentials.
• Whoever can read the latet-data volume can read the vault: it holds the key the vault is encrypted with. To keep the key elsewhere, give it as a Docker secret and set LATET_ROOT_KEY_FILE to its path.

Behind a proxy

When the providers are reached through a proxy of your network, set HTTPS_PROXY (and NO_PROXY) before docker compose up; compose.yaml passes them on. If the proxy inspects TLS with a certificate of its own, mount that certificate in the container and set SSL_CERT_FILE to it.

Undo

This takes Claude Code’s settings out (the same for codex and opencode, with their files):

docker exec -i latet latet agents merge claude --remove < ~/.claude/settings.json > ~/.claude/settings.json.new \
  && mv ~/.claude/settings.json.new ~/.claude/settings.json

claude plugin uninstall latet@latet removes the plugin. Then docker compose down; docker compose down -v also deletes the volumes, and with them the vault.