Models
Models folder
Permissions
What the assistant is allowed to do
Internet access (weather & search)
Lets it check real weather (Open-Meteo) and search Wikipedia for general information. No other network access is granted.
Code execution (Python sandbox)
Lets it run Python snippets on this machine via the sandboxed run_python tool. Also requires the server to be started with --enable-python-tool.
File access (attachments & workspace)
Lets you attach text/code files and paste long snippets into the chat, and lets run_python read/write files in its per-session workspace folder.
Microphone (voice input) — temporarily disabled
Lets you record and transcribe speech into the composer. Nothing is sent anywhere until you press record. The mic button is hidden for now regardless of this toggle - see the known issues doc for the transcription reliability bug being worked on.
Settings
Model
Sampling
Higher = more varied/creative. Lower = more focused and deterministic.
Keeps only the most likely next words that add up to this probability mass.
Limits each next-word choice to the k most likely candidates.
Above 1.0 discourages the model from repeating the same words or phrases.
Hard cap on how long a single reply is allowed to get.
Start a conversation
Everything runs locally through this server's own engine — pick a model and tune sampling from Settings, or just start typing to use the default.
No workspace — code/file tools off
Choose a workspace folder