Agent

The <perspective-viewer> chat sidebar, configured through a simple provider form. The viewer's embedded agent (viewer.agentConfig({...})) drives the view through its tool surface — schema inspection, view configuration, plugin selection and ExprTK expression authoring — against any of these providers:

• OpenAI-compatible — any local or remote server speaking /chat/completions, e.g. LM Studio's developer server (http://localhost:1234/v1, enable CORS in its server settings), Ollama (http://localhost:11434/v1, set OLLAMA_ORIGINS), llama.cpp or vLLM. Tool-calling quality is model-dependent — Qwen and Llama 3.1+ instruct models work well.
• Anthropic — claude-opus-5 by default; requires an API key, which is sent directly from the browser (Anthropic's CORS opt-in) — suitable for local development; use a baseUrl proxy for anything shared.
• Gemini — gemini-2.5-flash by default; requires an API key, sent directly from the browser (the Gemini API is CORS-enabled) — same local-development caveat.
• WebLLM — runs the model in this tab on WebGPU, so no prompt or data leaves the machine and no key is needed. Selecting it lazily imports @mlc-ai/web-llm and downloads model weights (gigabytes on a cold cache; the browser caches them afterwards). Only the Hermes family is offered — see the caveat below.

CORS-friendly hosted services that speak the OpenAI wire format (e.g. Groq at https://api.groq.com/openai/v1, or OpenRouter at https://openrouter.ai/api/v1) work through the OpenAI-compatible provider — just set the base URL and key.