Providers

elanous can use several model providers. Pick one as the default with elanous config set llm.provider <id>; you can still choose a different one per run or per role.

Provider	How you pay	What you need	Set up
openai-codex	ChatGPT subscription	a ChatGPT plan	elanous login openai-codex
grok	xAI subscription or API credit	XAI_API_KEY or a signed-in grok CLI	put the key in ~/.cache/xai_api_key
openrouter	prepaid credit	OPENROUTER_API_KEY	put the key in ~/.cache/openrouter_api_key
anthropic	API credit	ANTHROPIC_API_KEY	put the key in ~/.cache/anthropic_api_key
gemini	API credit	GEMINI_API_KEY	put the key in ~/.cache/gemini_api_key
local	free	an OpenAI-compatible local server	elanous local setup

elanous doctor shows which of these resolve on your machine. elanous usage shows what is left on each account.

What auto does

With llm.provider left on auto (the default), elanous uses a ChatGPT sign-in first. When one ChatGPT account runs low, it moves to another signed-in account; when none is left, it follows llm.fallbackChain (default: Codex accounts, then Grok), and after that any other provider that has a key. Set llm.fallbackChain to ["codex-rotate"] if you never want it to move to Grok.

Kimi, GLM and Qwen through OpenRouter

One OpenRouter key gives access to many vendors' models. Model ids look like openrouter/<vendor>/<model>.

elanous registry discover --source openrouter            # preview the model list (no key needed)
elanous registry discover --source openrouter --write    # add it to your local catalog

Use it for the part of a run that writes code, and keep your default for the rest:

elanous harness say --child-llm-provider openrouter --child-llm-model openrouter/moonshotai/kimi-k3 "<sentence>"

Or make it the default: elanous config set llm.provider openrouter.

Choosing a model by role

elanous harness plan --role-llm <role>=<provider>[/<tier>] "<sentence>" picks a model for one role in one run; the roleLlm setting makes it permanent. Tiers are budget, balanced, better, best, loaded; elanous tier shows the model behind each.

Did it actually use the model I chose?

Check after a run instead of assuming:

elanous logs --since 1h --grep <model id>

Enterprise endpoints (Bedrock, Vertex)

Bedrock and Vertex are not supported today — they are deliberately blocked. The agent driver scrubs CLAUDE_CODE_USE_BEDROCK and CLAUDE_CODE_USE_VERTEX from the child environment so that child agents run on subscription OAuth rather than on metered API credentials.