Great point on vouch decay — this is exactly the kind of thing we've been thinking about. Right now vouches are permanent, but adding TTL + automatic re-validation is on the roadmap. The idea is agents would need to periodically re-verify before trust propagates, and stale vouches would naturally lose weight in the trust score calculation.
Your approach with deterministic fingerprints and multi-validator policies is really interesting too. We're seeing similar patterns emerge — the "multiple independent validators" model maps well to how Joy's vouching works, where trust score increases with diverse, independent vouches rather than volume from a single source.
Appreciate the feedback. If you want to poke around, the MCP endpoint is live at https://joy-connect.fly.dev/mcp — any MCP-compatible agent can connect directly.
Quick update: We've rebranded to AutropicAI to unify under the Autropic company brand. Same API at tryautropic.com, same functionality, just a cleaner name that reflects our focus on AI efficiency.
Also just shipped our first API calls - the system is live and working! 860 models indexed, instant recommendations.
Next up: building an MCP server so Claude and other agents can use this automatically.
Hey HN! Built this because I kept overpaying for LLM API calls.
There are 800+ models with prices ranging from $0.01 to $60 per million tokens. Most devs just pick GPT-4 and move on.
SavvyLLM analyzes your task and returns the cheapest capable model. Unlike OpenRouter (5.5% proxy fee), we're purely advisory — you call providers directly with your own keys.