This project runs models on your own machine and talks to exactly one external service: the Hugging Face Hub, and only when you ask it to download something.
| when | what | where |
|---|---|---|
you press Download, or run llm-fetch.ps1 |
the repo id and file name you asked for | huggingface.co |
you press HF token..., or run hf auth login |
the token itself, as a request header | huggingface.co |
| everything else, always | nothing | - |
The model serves on 127.0.0.1 only (--host 127.0.0.1 is in every preset), so no other machine on
the network can reach it, and there is no telemetry, update check, or call home anywhere in bin\.
- Stored at
%USERPROFILE%\.cache\huggingface\token, which is where every Hugging Face tool looks. That is deliberate: the token is not this app's private file, it is the machine's. - Never passed on a command line (it travels as the
HF_TOKENenvironment variable) and never written to a log. - Public models need no token at all. A token buys you gated repos and higher rate limits.
- Revoke it at
huggingface.co/settings/tokensif it ever lands somewhere you did not intend.
No model weights, no engine binaries (the downloader fetches them), no credentials, no .env, and
.gitignore refuses *.env, token, stored_tokens, *credential*, *.pem, *.key.
Open a normal issue for anything that is not a vulnerability. If it is one - the launcher starting a
process it was not asked to, the downloader writing outside models\, a token reaching a log - say
so in an issue titled security: without reproducing the secret.