There is no server to send your code to.
Most privacy pages ask you to trust a policy. This one asks you to run a packet capture, because the claim is a property of the build rather than a promise about our conduct.
Verify it in four steps.
You do not have to take our word for any of this, and you shouldn't.
- Install Latorium and run Ollama locally.
- Start a packet capture, or set a default-deny outbound firewall rule with one exception for your local inference port.
- Use it normally — retrieval, drafting, completion, the task board.
- Watch nothing leave the host.
Every network call in the extension targets the inference URL you configure, which defaults to loopback. There is no analytics endpoint, no crash reporter, and no account system to attach any of it to.
To be exact about the one piece that isn't the extension: buying a licence goes through a payment provider, and a key is issued and emailed to you. The extension itself never talks to that — it verifies the key offline against an embedded public key, which is why it still works air-gapped.
What is written, and where.
Everything Latorium records stays inside your workspace, in files you can read and delete.
- Audit log
- your workspace
- Task board
- your workspace
- Saved lessons
- your workspace
- Licence key
- verified offline
Session token counters are read from the logs Claude Code and Codex already write on your machine. Only the counters — never prompt or response content.
The part we won't overclaim.
Latorium can keep your code on your machine. It cannot make the agent you point it at do the same.
Claude Code and Codex are cloud products. Whatever they read still goes to their provider under their terms — Latorium's effect is that they read less, not that they read locally. If you need nothing to leave the host at all, run the local model on its own, without a cloud agent attached.
We'd rather write that down than let the privacy claim quietly imply something it doesn't cover.