AI providers and cost
Choose a cloud or local model, keep token costs low, and use a stronger model only for planning.
Opsaido works with every major AI service and with local models. Your systems’ passwords and keys are encrypted and never sent to the AI.
Providers
| Type | Providers | Good for |
|---|---|---|
| Cloud | Anthropic Claude, OpenAI, Azure OpenAI, Google Gemini, Amazon Bedrock, Mistral AI, xAI Grok, DeepSeek | Best answers and planning |
| Fast / low-cost cloud | Groq, Cerebras, OpenRouter, Together AI, Fireworks AI, DeepInfra, NVIDIA NIM, Hugging Face, Moonshot Kimi, Alibaba Qwen | Cheap everyday questions |
| Local | Ollama, LM Studio, vLLM, LocalAI, llama.cpp server, Jan | Fully private, no token costs |
| Anything else | Any OpenAI-compatible endpoint | Your own gateway or proxy |
Add one in Settings → AI providers → Add. Each preset fills in the address, and the model list loads live.
Cost settings
Every new provider starts with cost-saving defaults. Change them under Cost & response.
| Setting | Default | What it does |
|---|---|---|
| Chat memory | 24,000 tokens | How much of the earlier conversation goes with each request. Lower is cheaper. |
| Max answer length | 6,000 tokens | A cap on one reply. You pay only for what’s used. |
| Reasoning effort | Low | How hard reasoning models think. Low suits most work; thinking is billed as output. |
| Planning model | — | A stronger model used only for dry runs and plans. Everyday questions stay on the cheaper one. |
| Temperature / Top P | Provider default | Only for models that support them. |
A good mix: a low-cost model for everyday questions, plus a stronger planning model for dry runs. You get careful plans without paying top prices for every “how’s the disk?”.
How Opsaido keeps requests small
| What | Effect |
|---|---|
| Only relevant tools are sent with each question | About two thirds fewer tokens per step |
| The tool list stays stable within a chat | The provider’s prompt cache keeps working (cached input costs a fraction) |
| Old tool output is shortened in chat memory | Long chats don’t grow without limit |
| Fleet tools check a whole cluster or many hosts in one call | One step instead of dozens |
What each chat costs
Each chat’s Usage button shows what that conversation has used so far, for example 48.2k tokens · $0.06. It updates live during a run, and the tooltip breaks it down into input, output and cached tokens. The cost appears once you’ve set prices for the provider. To hide it for everyone, switch off Token meter in chat in Settings → AI providers.
Budgets
Set prices and a monthly budget for each provider. The live panel shows tokens and cost, and you get an alert at 80% and 100% of the budget.
Reply language
The assistant answers in the language each person writes in. To use one language for everyone, pick it in Settings → General → Reply language. Commands, code, file paths and names are never translated.