Anthropic

Claude as the bot's brain for text replies, and (through the cascade) for voice.

Key: create one at console.anthropic.com (Settings, API keys). Anthropic keys start with sk-ant-.

What it's used for

Use	Supported
Text replies (the brain)	Yes
Images and camera frames	Yes, every Claude model
Tools (built-in, HTTP, MCP, page tools)	Yes
Voice	Through the cascade: pair Claude with speech to text from OpenAI or ElevenLabs and a voice from ElevenLabs, OpenAI or Gemini
Describing camera frames for ElevenLabs Agents	Yes (vision.proxy)

Models

The model list comes from your key. A new bot starts with claude-opus-5-5 when the key has it (then claude-sonnet-5-5, claude-opus-5, claude-sonnet-5, else the first listed).

• brain.effort is sent as Claude's effort setting (minimal counts as low); models without it ignore it.
• brain.temperature is only sent to older models that still take it.
• The system prompt is cached between turns, which makes long conversations cheaper and faster.
• Replies stream token by token, and tool calls show as "Using ..." in the chat while they run.

Refusals

On the official API (not when PROVIDER_BASE_URL_ANTHROPIC points elsewhere), the server turns on the API's server-side fallback for refusals on the models that support it (claude-opus-5-5, claude-opus-5, claude-sonnet-5-5 and claude-fable-5-1). If Claude still declines, the visitor gets a short "Sorry, I can't help with that one" and the face looks concerned; the conversation carries on.