Ollama models

Cluster Build can use an Ollama server that is reachable over HTTP. Configure one named provider in ~/.cluster/config.toml. The default provider only sets the startup selection; /model also lists Cluster models and every reachable configured Ollama provider:

default_model_provider = "gpuserver"
default_model = "qwen3.5:9b"

[model_providers.gpuserver]
kind = "ollama"
base_url = "http://gpuserver:11434"

For Ollama running on the same computer, a typical provider is:

[model_providers.local]
kind = "ollama"
base_url = "http://127.0.0.1:11434"

base_url is the host root; Cluster Build appends /v1. Ollama providers are keyless, so starting the TUI does not require a Cluster login and requests do not carry Cluster credentials. The existing /model picker lists the models returned by that server's /v1/models endpoint.

Ollama must listen on an address reachable from the client. Keep port 11434 limited to the trusted private network because Ollama does not authenticate these requests.

Context length

Coding-agent prompts and tool schemas require a much larger context than Ollama's 4K default on GPUs with less than 24 GiB of VRAM. Configure at least 32K before starting Ollama, then confirm the active allocation with ollama ps:

[Service]
Environment="OLLAMA_CONTEXT_LENGTH=32768"

Known limits

Cluster Build disables Ollama thinking for this first provider slice so a tool follow-up always produces a visible assistant answer. Reasoning controls are not exposed for Ollama yet.

Switching providers

Selecting a model from another provider clears history and starts a new session before changing request routing. Switching to Cluster authenticates when needed; switching to Ollama remains keyless. Resuming a local session only requires its named provider to remain configured.

Next steps

• Advanced configuration — the rest of config.toml, including the [model_providers] table.
• Interactive sessions — the /model picker and other slash commands.