point DGC at Ollama, llama.cpp, vLLM, or a cloud host

DGC talks to any supported native or OpenAI-compatible endpoint. Pick one with /connect (or dgc setup on first run), then /model to choose the served model.

• Ollama — /connect ollama uses http://localhost:11434/v1. DGC auto-detects Ollama and speaks its native chat API (which round-trips the model's own thinking).
• llama.cpp — run llama-server, then /connect llamacpp (http://localhost:8080/v1, no real key needed). This is the OpenAI-compatible /v1 server built into llama.cpp.
• vLLM / SGLang — /connect vllm (http://localhost:8000/v1). The server renders the chat template, so DGC sends the reasoning switch it understands.
• LM Studio — /connect lmstudio (http://localhost:1234/v1).

unsloth is not a serving runtime — its GGUF/quantized models are served through llama.cpp (llama-server), vLLM, or Ollama. Start one of those with your unsloth model, then pick that runtime's preset. /think <level> reaches Qwen3-family templates (which read the effort from inside chat_template_kwargs) automatically.

/connect http://<host>:<port> points DGC at any other OpenAI-compatible server. If you enter a bare host, DGC appends the /v1 chat-completions path for you and prints a notice (Anthropic and native Ollama URLs are left as-is). Models are auto-discovered from the endpoint's /v1/models, so /model lists what the host actually serves.

/connect openai | anthropic | openrouter | groq | deepseek | together | mistral prompt for the provider's key and use its native auth contract. See Subscriptions to instead run your own Claude/Codex/Qwen/Kimi/Copilot plan through its official CLI.