Providers and authentication
LiteLLM unifies 100+ backends behind one interface. On top, Okami routes by capability, cost and availability, with a fallback chain. The posture is subscription-on…
Catalog
| provider | models | auth · tier |
|---|---|---|
| claude | Opus · Sonnet · Haiku | claude CLI · strong |
| codex | GPT (Codex family) | OAuth device · strong |
| minimax | MiniMax (long context) | OAuth token · weak |
| mimo | MiMo | API key · weak |
| lmstudio | any local model (default) | OpenAI-compat · local |
Tiers and capability profile
Each (provider × model) pair carries a capability profile derived from the tier. It tells the harness how much to decompose, which tool-call mode to use and when to ensemble. The learning loop auto-tunes the profile by measuring each model's failure modes.
| knob | weak / local | strong |
|---|---|---|
| Decomposition | fully-specified micro-steps | large chunks |
| Tool-calls | JSON/GBNF constrained → ReAct | native |
| Context / tools | lean window, small subset | broad |
| Critical decisions | vote N / escalate early | single-shot |
Authentication per provider
params claude CLI Uses the claude CLI (already logged in). Don't use a pay-as-you-go ANTHROPIC_API_KEY — the ToS restricts it. codex OAuth device okami login codex (OpenAI native device flow; enable Device Code in ChatGPT). minimax OAuth device okami login minimax (OAuth device flow, token plan). mimo API key Set MIMO_API_KEY in .env. lmstudio OpenAI-compat Point api_base at your local server (default http://localhost:4480/v1).
Chained fallback
Each provider declares a fallback. If codex fails (529 / timeout / empty response), the router tries claude → minimax → lmstudio, in order, without you noticing. That's what kills vendor lock-in.
providers:
codex:
model: openai-codex/<model>
fallback: [claude, minimax, lmstudio]
tier: strong
context_window: 256000
reasoning_effort: high # minimal | low | medium | high