Loading…
Loading…
Local Mode
Point AGI at a model server you already run. The work happens on your hardware, with no AGI account and no meter, and the only route out of Local mode is one you read before you send it.
Surfaces
Setup
Nothing is installed on your behalf. Start the runtime, pull or load a model the way you normally would, then let AGI discover what is there and name the provider explicitly.
# an Ollama server is already running $ ollama pull <your-model> $ agi models scan Local model servers ollama running http://localhost:11434 models: - <your-model> lmstudio not running http://localhost:1234/v1 $ agi --provider ollama --model <model> $ agi --provider lmstudio --model <model>
The trust boundary
A session carries its own privacy authority — local, byok, or managed — written down beside its transcript, and the runtime re-reads it before each request it builds. When the selected model routes somewhere that authority does not cover, the send stops there and the error names the one command that would carry the work across properly.
That command forks a second session which starts with no history at all. It runs a secret scan over the messages you picked, shows you the exact payload with counts of what was included, excluded and truncated, and completes only when you send that reviewed draft back word for word. Attached files are never added for you. The Local session it forked from keeps its own authority and its own transcript, unchanged.
Local Mode
Desktop carries the widest set of local runtimes. BYOK is the reviewed way out on the days a local model is not enough.
Public launch: date to be announced