LLM providers

AI columns, research columns, agent columns and chat all go through one LLM client with automatic failover. You need at least one key; several providers have free tiers.

Supported providers

id	Key	Default model	Base URL override
`anthropic`	`ANTHROPIC_API_KEY`	`claude-opus-4-8`	`ANTHROPIC_BASE_URL`
`cerebras`	`CEREBRAS_API_KEY`	`llama-3.3-70b`	`CEREBRAS_BASE_URL`
`groq`	`GROQ_API_KEY`	`llama-3.3-70b-versatile`	`GROQ_BASE_URL`
`sambanova`	`SAMBANOVA_API_KEY`	`Meta-Llama-3.3-70B-Instruct`	`SAMBANOVA_BASE_URL`
`nvidia`	`NVIDIA_API_KEY`	`meta/llama-3.3-70b-instruct`	`NVIDIA_BASE_URL`
`mistral`	`MISTRAL_API_KEY`	`mistral-small-latest`	`MISTRAL_BASE_URL`
`openrouter`	`OPENROUTER_API_KEY`	`meta-llama/llama-3.3-70b-instruct:free`	`OPENROUTER_BASE_URL`
`github_models`	`GITHUB_MODELS_API_KEY`	`gpt-4o`	`GITHUB_MODELS_BASE_URL`
`siliconflow`	`SILICONFLOW_API_KEY`	`Qwen/Qwen3-8B`	`SILICONFLOW_BASE_URL`

Each has a <PROVIDER>_MODEL variable to change the model. The Settings UI additionally lists Google AI (GOOGLE_AI_API_KEY, gemini-2.5-flash), Cohere, Cloudflare Workers AI and Hugging Face with their free-tier notes. Anthropic uses the native Messages API; the others are OpenAI-compatible.

Which provider runs

LLM_DEFAULT_PROVIDER, when set, always wins.
Otherwise, if an Anthropic key is present and YUPCHA_CLOUD is truthy, Claude becomes the default. Self-hosted installs stay opt-in.
Otherwise cerebras.

Failover order after the default is cerebras, groq, sambanova, nvidia, mistral, openrouter, github_models, siliconflow, anthropic. Anthropic sits last so the free tier is never displaced as a fallback. A call tries up to three providers, with exponential backoff and jitter on rate limits.

Keys saved in the Settings UI override environment variables.

Tuning

Setting	Default	Meaning
`LLM_PROMPT_CACHE`	on	Cache identical prompts
`LLM_HTTP_TIMEOUT`	60	Seconds per request
`CHAT_MAX_TOOL_ROUNDS`	8	ReAct tool rounds per chat turn
`RESEARCH_NATIVE_TOOLS`	off	Use Claude native tool-use for research columns when Anthropic is the default
`RESEARCH_CELL_BUDGET_USD`	0.05	Per-cell budget for research columns

Prompt injection

Fetched web content is fenced as untrusted data before it reaches a model, and research columns use in-house search/fetch tools rather than vendor server-side browsing so every fetch passes the SSRF guard and the fence. See Security model.