Hosted models

bnw inference register <CAPABILITY_ID> --preset openrouter --runtime-model <vendor/model> sets up the OpenAI-compatible adapter for OpenRouter. --preset huggingface does the same for Hugging Face inference providers, with models named org/model[:provider]. Each preset fills in the endpoint, the usual key secret (OPENROUTER_API_KEY or HF_TOKEN), remote egress, and the router’s headers. From its public model list, it also works out whether the model takes tools. bnw inference models openrouter|huggingface browses those lists. bnw share model <MODEL> --via openrouter|huggingface, or the web client’s Share a model card, shares one in a single step. The Hugging Face MCP server is set up with --auth bearer --token-env HF_TOKEN (see above). INFERENCE.md has the details, including how OpenRouter’s reported cost is kept for the provider only.