What leaves your machine
Saathi is a program you run. What leaves your machine depends entirely on which mode you are in, and the default mode sends nothing anywhere.
The three modes, in full
localDefault
An OpenAI-compatible server on your own machine: Ollama, LM Studio, llama.cpp. With nothing configured at all, this is what you get.
- Keys
- none
- Account
- none
- Leaves device
- no
your own key
OpenAI, Anthropic, or Sarvam AI. The key sits in ~/.saathi/shell.json on your machine and the request goes straight to that provider. Saathi’s servers are not in the path.
- Keys
- yours, on your machine
- Account
- none
- Leaves device
- yes
hosted
For people who would rather not run or configure anything. It runs the same open-source code, with accounts switched on by configuration rather than by a private fork.
- Keys
- we hold them
- Account
- required
- Leaves device
- yes
There is no silent fallback
If a local model is not reachable, Saathi says so and stops. It never quietly upgrades to sending your words to a cloud provider instead. Both the macOS and the Windows client have tests pinning that behaviour, and those tests run on every change.
Your key stays yours
In your-own-key mode the key lives in ~/.saathi/shell.json on your machine and the request goes straight to that provider. Saathi’s servers are not in the path and never see it.
The config file is locked down on Unix, and not yet on Windows
On macOS and Linux the file holding your key is chmod 0600. The Windows equivalent is an ACL, and it is stubbed rather than implemented. It is written down as a comment in the code rather than silently skipped, and it has to land before anything writes a real token. Until then, treat that file as readable by anything running as you.
This website collects nothing
No analytics, no cookies, no third-party scripts, no email capture, no fonts fetched from anyone else’s server. It is a folder of static files.
The hosted backend, if you use it
The hosted mode holds provider keys on Saathi’s side, which is the whole point of it, and it requires an account. It runs the same open-source code you can read and deploy yourself.