Solution
Claude on Rails, done properly.
Anthropic prompt caching, tool use and streaming — exposed through the same Ruby DSL you already use for OpenAI.
One switch to change providers
class Concierge < RailsAgents::Base
model "claude-3.5-sonnet" # was "gpt-4o"
cache :system_prompt # anthropic prompt caching
endProvider-aware defaults
- Prompt caching flag on/off from the DSL
- Tool schemas translated automatically
- Cost dashboards per provider