LLM Providers

The seven model providers Minara supports, Auto on the Minara catalog, API-key and OAuth paths, billing risks, and how to pick one.

Minara needs one LLM provider to reason. It supports seven provider families. Most offer two ways to sign in: an API key (you pay per token) or OAuth (you reuse a subscription you already pay for). A couple are single-path: Ollama Cloud is API-key only, and Minara, the built-in platform account, is OAuth only. This page explains the tradeoffs, the billing risks, and how to choose. For the actual login commands see Installation; for the key names see Environment Variables.

At a glance

Provider	API-key path	OAuth path	OAuth runs on	Models on OAuth	Billing
Minara	— (none)	`minara-oauth`	Your Minara account	Auto plus the models the Minara gateway serves	Billed through your Minara account.
Anthropic	`anthropic-api-key`	`anthropic-oauth`	Claude Max plus extra usage credits	Claude models	API key: pay per token. OAuth: spends purchased extra-usage credits.
OpenAI	`openai-api-key`	`openai-oauth` (Codex)	A ChatGPT subscription, through the Codex backend	Codex-supported models only	API key: pay per token. OAuth: included in your ChatGPT plan.
OpenRouter	`openrouter-api-key`	`openrouter-pkce`	An OpenRouter account balance	300+ models, all vendors	Both paths: pay from OpenRouter credit.
Kimi	`kimi-api-key`	(none)	(none)	(none)	Pay per token through your Kimi API account.
xAI (Grok)	`xai-api-key`	`xai-oauth`	A SuperGrok or X Premium+ subscription	Grok models	API key: pay per token. OAuth: included in your X subscription.
Ollama Cloud	`ollama`	— (none)	—	—	Free tier with limits, plus paid Turbo. No per-token bill Minara meters.

Choosing a path: API key vs OAuth

An API key is a usage meter. You pay for exactly what you use, and the provider's full model lineup is available. This path works for everyone.

OAuth reuses a subscription you already bought. You do not get a separate per-token bill. In exchange, only some models work and there are plan limits. Each provider's OAuth path has its own catch, covered below.

Choose a model and reasoning level

Use the model control in the Chat composer to choose a model and its reasoning level. The model and reasoning options open in matching nested menus, with a consistent position and interaction pattern.

Your choice belongs to the current Chat session. Minara uses it for the next message and keeps it for later turns in that chat. It does not replace a response already in progress or change the deployment-wide default.

On the Minara provider, start with Auto. It scores each message and picks a catalog model for that turn. The section below covers the strategies. You can still pin a named model in the composer. The session choice takes priority for later messages in the current chat.

Minara

Minara is the built-in platform account. Sign in with minara-oauth. Usage bills through your Minara account. The catalog is whatever the Minara gateway currently serves.

Auto

Auto is the recommended model on Minara. Choose it once in the Chat composer. Minara scores the next message and routes it to a catalog model that fits the turn.

A quote check stays on a fast, inexpensive model. A market memo, an automation, a live order path, or a coding change steps up. You keep a snappy desk for the small stuff and a precise one for the work that moves money or ships software.

That mix is what Minara is for: financial analysis, scheduled and agent-run workflows, trade execution, and programming. Leave Auto on and one chat can do all of them. The router spends more only when the turn earns it.

Three strategies sit under Auto in Settings (Provider & Model) after Auto is selected. A saved user setting wins over MINARA_AUTO_MODEL_STRATEGY. Background title and compaction models stay on the Background task models page.

Strategy	Use it when	Default mapping
balanced (default)	You want one setting that stays quick on light turns and spends when the score rises.	DeepSeek V4 Flash on SIMPLE/MEDIUM (Claude Sonnet 5 if the turn needs vision or PDF). GPT-5.6 Terra on COMPLEX. GPT-5.6 Sol on REASONING.
economy	You want to stretch credits.	Same Flash path on light turns. GLM 5.2 on COMPLEX. Kimi K3 on REASONING.
quality	You want the strongest catalog pick at every difficulty.	Claude Sonnet 5 on light turns. GPT-5.6 Terra on COMPLEX. Claude Opus 5 on REASONING.

Pin a named catalog model any time you want one model for the rest of the session. Auto stays in the picker for the next chat.

Anthropic (Claude)

Anthropic gives you Claude.

anthropic-api-key sends requests to api.anthropic.com. Get a key from console.anthropic.com and pay per token. Every Claude model is available.

anthropic-oauth reuses your Claude Max subscription. Minara signs in the way the Claude Code tool does, so Anthropic treats your requests as Claude Code activity.

Anthropic OAuth needs Claude Max plus extra usage credits

minara auth login anthropic signs in the way Claude Code does. Anthropic bills those requests as Claude Code activity. This path works only on a Claude Max plan. It spends the "extra usage" credits you buy on top of Max. It never touches your base Max allowance. With no extra credits, requests fail even on a valid Max plan. Claude Pro cannot use this path. On Pro, or to pay per token, set ANTHROPIC_API_KEY instead.

OpenAI

OpenAI gives you two ways to sign in, and they are very different.

openai-api-key sends requests to api.openai.com. Get a key from platform.openai.com and pay per token. The full OpenAI lineup is available, including gpt-5.5-pro, gpt-4.1, and o3.

openai-oauth is Codex OAuth. Minara signs in through the OpenAI Codex CLI device flow and presents itself as the Codex CLI. Your requests go to OpenAI's first-party Codex backend at chatgpt.com/backend-api/codex, not to api.openai.com, and run on your ChatGPT subscription.

OpenAI OAuth is Codex OAuth, not the OpenAI platform API

minara auth login openai signs in as the OpenAI Codex CLI. Your requests run on your ChatGPT subscription through OpenAI's Codex backend. Only Codex-supported models work on this path. Models like gpt-5.5-pro, gpt-4.1, and o3 are not available here. To use the full OpenAI lineup, sign in with an OpenAI API key instead.

The ChatGPT Codex backend serves a rotating allow-list. The live catalog in Settings is the source of truth. Current-generation models include:

gpt-5.6-terra (default)
gpt-5.6-sol
gpt-5.6-luna

Older *-codex slugs such as gpt-5.3-codex return HTTP 400 once the backend retires them.

In Settings, the model picker shows the full OpenAI lineup at all times. Models your connected accounts cannot run appear greyed out with a short reason. On ChatGPT sign-in, the non-Codex models are greyed until you connect an OpenAI API key. If you have both a ChatGPT sign-in and an API key connected, you can pick any model. Choosing one that needs the other account switches your active connection for you.

OpenRouter

OpenRouter is one front door to more than 300 models from many vendors (Anthropic, OpenAI, Google, Meta, and more). Both sign-in paths bill from the same OpenRouter credit balance.

openrouter-api-key uses a key from openrouter.ai. openrouter-pkce signs you in through your browser, with no key to copy. Either way Minara attaches attribution headers so your usage is credited to Minara on the OpenRouter dashboard.

OpenRouter always draws from your OpenRouter balance

There is no subscription path here. Both the API key and the browser sign-in spend credit you have loaded into your OpenRouter account. Top up your balance before a long session. When the balance hits zero, requests stop until you add more.

Kimi

Kimi is an API-key-only provider backed by the official Kimi API platform. Create a key at platform.kimi.ai, then run minara auth login kimi --api-key <key> or set KIMI_API_KEY.

kimi-api-key sends OpenAI-compatible Chat Completions requests to https://api.moonshot.ai/v1. The live model list comes from Kimi's official /v1/models endpoint; kimi-k2.6 is the cold-start default. Streaming, tool use, structured output, and K2.6 reasoning-content replay are supported. When a turn must force one tool, Minara disables K2.6 thinking for that call because the Kimi API only allows automatic or disabled tool choice while thinking is enabled.

xAI (Grok)

xAI gives you Grok.

xai-api-key uses a key from the xAI console and bills per token at api.x.ai. xai-oauth signs in through your browser and reuses your SuperGrok or X Premium+ subscription.

xAI OAuth needs a SuperGrok or X Premium+ subscription

This path reuses the Grok access that comes with those plans. A free X account does not work. Without one of these subscriptions, use an xAI API key instead.

Ollama Cloud

Ollama Cloud runs open models (gpt-oss, DeepSeek, Qwen3-Coder, Kimi K2, and more) on Ollama's hosted servers, so you do not need a local GPU. It has one sign-in path: an API key.

ollama sends requests to https://ollama.com over an OpenAI-compatible endpoint. Create a key at ollama.com/settings/keys, then run minara auth login ollama --api-key <key> (or set OLLAMA_API_KEY). To point at a self-hosted or local Ollama instead, set OLLAMA_BASE_URL (for example http://localhost:11434) or pass --base-url.

Ollama Cloud has free-tier limits and a paid Turbo tier

The free tier caps how much you can run. Heavy use needs the paid Turbo plan. When you hit a limit, requests fail until the window resets or you upgrade. There is no per-token bill Minara meters, so the cost panel shows zero for this provider.

Tool calling, vision, and JSON output depend on the specific cloud model you pick. The larger instruction-tuned models (for example gpt-oss:120b) handle tools well; smaller models may not.

Which should I pick?

Your situation	Pick this	Why
You signed in with Minara and want one model that scales with the work	Auto on the Minara provider	Routes each message. Fast when the turn is light, stronger when the work is analysis, automation, trading, or code.
You pay for Claude Max and bought extra usage credits	Anthropic OAuth (`anthropic-oauth`)	Reuses credits you already own, no new bill.
You want Claude but only have Pro, or no subscription	Anthropic API key (`anthropic-api-key`)	Pro cannot use the OAuth path. Pay per token instead.
You pay for ChatGPT and only need coding-grade models	OpenAI OAuth (`openai-oauth`)	Runs on your ChatGPT plan, no extra bill. Codex models only.
You need `gpt-5.5-pro`, `gpt-4.1`, `o3`, or the full lineup	OpenAI API key (`openai-api-key`)	OAuth cannot reach these. The API key can.
You want to try many models from different vendors	OpenRouter	One account, 300+ models.
You want native Kimi models and already have a Kimi API account	Kimi (`kimi-api-key`)	Direct official endpoint, live Kimi model catalog, no aggregator hop.
You want Grok and pay for SuperGrok or Premium+	xAI OAuth (`xai-oauth`)	Reuses your subscription.
You want open-weight models behind one hosted key, no local GPU	Ollama Cloud (`ollama`)	One key, no per-token bill, runs gpt-oss / DeepSeek / Qwen.
You are not sure and want the simplest start	Any API key	Works for everyone, every model available, you pay only for what you use.

Troubleshooting

Error: out_of_credits or insufficient credit

Your provider balance ran out. For OpenRouter, top up at openrouter.ai. For Anthropic OAuth, buy more extra usage credits on your Claude Max plan. For an API-key path, add funds or raise the spend limit in the provider console.

A model returns 404 or model not found on OpenAI OAuth

You asked for a model the Codex backend does not serve. Pick a model from the live Codex list such as gpt-5.6-terra, or sign in with an OpenAI API key to reach the full lineup.

Sign-in is rejected or the session expired

Your saved credential is invalid or out of date. Run the matching minara auth login <provider> command again to refresh it. For Anthropic OAuth, confirm you are on Claude Max, not Pro. For OpenAI OAuth, confirm your ChatGPT subscription is active.