Ollama, managed vLLM Metal, LiteLLM, cloud passthroughs, Open WebUI, backend, MCP/tool access, and trace hooks.

Open the full-size diagram.

Disabling a cloud provider in .env doesn't error — the model resolver silently produces zero catalog entries for it. Managed vLLM Metal is another local LiteLLM upstream, so LLM_PROVIDER_SOURCE=none does not require a cloud provider. Ollama models get two aliases (ollama/<name> and the bare name); either works. Tracing observes requests out-of-band and isn't part of the inference call path, so a tracing-backend outage doesn't affect completions.

• services/litellm/service.yml
• services/litellm/models.yaml
• services/ollama/service.yml
• services/vllm-metal/service.yml