Provider comparison
| Provider | Best for | Pricing model | Data handling |
|---|---|---|---|
| OpenRouter | Flexibility — access 200+ models through one API key | Pay per token, varies by model | Pass-through to underlying provider |
| Anthropic | Long-form reasoning, careful tool use, safety | Pay per token (input + output) | Anthropic API — US-based |
| OpenAI | Broad model range, strong vision capabilities | Pay per token or subscription | OpenAI API — US-based |
| Local (llama.cpp) | Maximum privacy — prompts never leave your machine | Free (your hardware) | Local only — nothing sent externally |
Step 1: Get an API key
For cloud providers, you need an API key:
- OpenRouter: openrouter.ai/keys — free credits on signup
- Anthropic: console.anthropic.com — requires account
- OpenAI: platform.openai.com/api-keys — requires account
Step 2: Configure in Hermes
Run the setup wizard:
hermes setup
Or edit ~/.hermes/config.yaml directly:
provider: openrouter model: anthropic/claude-sonnet-4
Store your API key in ~/.hermes/.env:
OPENROUTER_API_KEY=sk-or-v1-...
Step 3: Test your configuration
hermes > What model are you running on?
Hermes should identify its current model. If not, check your key and config.
Recommended models for different tasks
- Quick tasks, low cost: Claude Haiku, GPT-4o mini, Gemini Flash — sub-$1/M tokens
- Complex reasoning: Claude Sonnet 4, GPT-4o — $3-15/M tokens
- Vision tasks: GPT-4o, Claude models with vision support
- Code generation: Claude Sonnet 4, DeepSeek v3
Canadian data considerations
When using cloud providers, your prompts are processed on their infrastructure — typically in the US. For organizations handling personal information under PIPEDA:
- Review the provider's data processing terms
- Consider using local models for sensitive workflows
- Check if the provider uses your data for training (most API providers do not)
- Document which provider handles which types of tasks
Not legal advice. Review your obligations with qualified counsel before sending regulated data to any API.