llama-server (no Ollama)
Run 8gent Code on llama.cpp's llama-server. With it switched on, 8gent never contacts Ollama.
Update to get this
llama-server support needs a current release. Run 8gent update; see Install.
8gent can run on llama.cpp's llama-server instead of Ollama. Switched on, nothing in 8gent probes, lists or calls Ollama, so you do not need it installed. If you do nothing, 8gent uses Ollama as before.
Turn it on
-
Start
llama-serverwith a chat model:llama-server -m ./your-model.gguf --port 8080 -
Tell 8gent to use it:
export EIGHT_LOCAL_SERVER=llama-server 8gent --provider=llama-serverInside a session:
/provider llama-server.
On a machine with no saved provider, 8gent picks llama-server on its own once EIGHT_LOCAL_SERVER=llama-server is set and the server answers. If you have used another provider before, pass --provider=llama-server once so it wins over the saved one.
Settings
| Variable | What it does | Default |
|---|---|---|
EIGHT_LOCAL_SERVER | llama-server to use llama-server; ollama or unset for Ollama | ollama |
LLAMA_SERVER_URL | Where your llama-server is. A trailing /v1 is fine | http://127.0.0.1:8080 |
What changes
- The model list shows the model or models your llama-server serves.
- The welcome screen does not warn that Ollama is missing.
- Task routing is skipped: every turn goes to your llama-server model.
System One without Ollama
The System One judge can run in-process too. Put its GGUF file in ~/.8gent/models/decide/, or point at it with EIGHT_DECIDE_GGUF=/path/to/model.gguf. With llama-server on, each 8gent process loads its own copy; the shared judge needs Ollama. See System One.
Turn it off
Unset EIGHT_LOCAL_SERVER (or set it to ollama) and switch back with /provider.