How models work
8gent runs on a model on your own machine by default. Pick the engine, pick the model, and add hosted models only if you want them.
Local first
8gent needs a language model to think with. By default that model runs on your machine, through a local engine:
| Engine | On by default | Page |
|---|---|---|
| Ollama | Yes | Ollama |
| llama-server (llama.cpp) | When you set EIGHT_LOCAL_SERVER=llama-server | llama-server |
| LM Studio | Yes, if it is running | LM Studio and Apple Foundation |
| Apple Foundation | On macOS 26 or later on Apple Silicon | LM Studio and Apple Foundation |
Nothing here needs an account or an API key. Hosted models are opt-in.
Which model it uses
- First run. With nothing saved yet, 8gent looks at what your engines serve and picks from what it finds. It says what it found on the welcome screen.
- After that, your choice is saved in
~/.8gent/providers.json(activeProvider,activeModel) and used every time. Auto-detection does not run again while a choice is saved. - Switch any time:
| How | Scope |
|---|---|
/model, /model <name> | This tab, now. With no name, a list you scroll with the arrow keys |
/provider, /provider <name> | This tab, now |
8gent --provider=<name> --model=<name> | This launch, and saved |
models.tabs in ~/.8gent/settings.json | Default model per tab (Orchestrator, Engineer, QA) |
If you add a new engine later, pass --provider once so it wins over the saved choice.
Routing by task
The task router can send different kinds of request to different models. It sorts prompts into four categories: code, reasoning, simple, creative.
/router on
/router set code qwen3:14b
/router set simple qwen3.5:9b
/router test "rename this variable everywhere"
/router statusWith llama-server, routing is skipped: every turn goes to your llama-server model.
When a model fails
If a model does not answer, 8gent tries the next one in its failover chain. Chains are kept in ~/.8gent/failover.json. Hosted entries in a chain are only used when that provider has a key.
Helper models
Besides the chat model, a few small local models do specific jobs: the System One judge, vision, and memory embeddings. See Helper and judge models.