Which provider should I start with? A cost-first guide to all 21

[AgentiLoop]

"Which provider should I pick?" is the most common first question, so here's the short version organized by budget. All 21 are in Settings → Provider; you can configure several and Agent! will fall back to the next one on a 429, timeout or network failure.

Free

Provider	Notes
Apple Intelligence	On-device, zero tokens, nothing leaves the Mac. Used via the brain icon for triage, summaries and token compression rather than as a main provider.
Ollama / vLLM / LM Studio (local)	Free in the API-fee sense only. A usable 30B+ model wants an M2/M3/M4 Ultra Mac Studio with 64–128GB. Agent! now reads the model's real context window from the server instead of assuming 32K.
Ollama (cloud)	Free tier for hosted open models.

Cheap — best starting point for real work

Provider	Notes
Z.ai / BigModel	GLM-5.3. Fastest signup, the default model, pennies per million tokens. This is the recommended first key.
DeepSeek	Budget coding; Agent! shows cache-hit stats.
MiniMax	1M-token context.
A2Agent	DeepSeek, GLM, Kimi, MiniMax and Qwen through one OpenAI-compatible key at a fraction of official pricing.
Alibaba DashScope	Qwen 3.8 pay-as-you-go (sk- or sk-ws- keys).

Subscription (no per-token billing)

Provider	Notes
Claude (OAuth)	Use your existing Claude subscription. Also supports API-key billing, extended thinking and prompt caching. Best for long autonomous tasks.
Codex	OpenAI models through ChatGPT OAuth — no API key.
Qwen / Qwen Code	Alibaba Token Plan / Coding Plan (sk-sp- keys): qwen3.8-max, qwen3-coder-plus, GLM-5, Kimi K2.5, MiniMax-M2.5.

Paid, per-token

Provider	Notes
OpenAI	General purpose, tool calling, vision, reasoning_effort pass-through.
Google Gemini	Long context, vision; has a free tier.
Grok (xAI)	Real-time info.
Mistral / Mistral Vibe	Open-weight cloud, code, agent product.
Hugging Face	Open models, serverless or dedicated endpoints.

Routers (one key, many models)

Provider	Notes
OpenRouter	200+ models; Claude is routed through the Anthropic protocol so caching/thinking still work.
OrcaRouter	190+ models at provider cost; auto/fusion/fallback routing.
Requesty	300+ models with per-model pricing and capability metadata.

A practical setup

1. Z.ai (GLM-5.3) as primary — cheap enough to leave running.
2. Claude or Codex via OAuth as the fallback for hard, long-running refactors.
3. Apple Intelligence enabled for on-device compression.

Sub-agents can override the model per agent, so you can have a fast/cheap model doing research fan-out while the parent runs on something stronger.

What's your provider mix? Reply with what's working (and what isn't).