Models & providers

Two wire protocols cover everything — pick a provider during setup or re-run bwn init. Environment variables override stored keys, so CI just works.

Hot-swapping with /model

/model switches model and provider mid-session, and never claims success without proof: a one-token probe through the real request path validates the key, the model name, and the server before anything is saved. A missing API key is configured right in the prompt; a stopped Ollama gets install/serve/pull steps; a failed probe keeps your current model.

/model gpt-4o                        # provider inferred and validated
/model ollama/qwen2.5-coder          # checked against `ollama list` first
/model meta-llama/llama-3.3-70b      # any OpenRouter model, by org/model name
/model http://localhost:8000/v1 my-model   # any OpenAI-compatible server (vLLM, TGI, LiteLLM)

Custom endpoints work keyless (the vLLM default); if you configure CUSTOM_API_KEY, bwn refuses to send it over plain HTTP to anything but a loopback address.

Local models

The native Ollama path probes each model's real context window and sets num_ctx so long prompts aren't silently truncated. Tool calls emitted as tagged text by small coder models (Qwen2.5-Coder, Gemma) are parsed reliably — a 1.5B model actually writes files.

Vision

Image and video attachments are gated per model: every current Claude model accepts images; OpenAI-compatible and Ollama models are detected by name (gpt-4o/o3/gemini, llava, qwen-vl, gemma3, llama3.2-vision, moondream, …). Text-only models get an explicit notice instead of silently dropped attachments. See Multimodal input.

Keys live in ~/.buildwithnexus/.env.keys (0600).