Built in the open

These aren't side projects; they're exhibits A and B for one claim: the unglamorous layer everyone skips is the layer that decides whether your agents survive production.

mintkey: your agents never hold real credentials

A credential broker for AI agents: agents get short-lived, scoped tokens; a proxy injects the real secret; every call lands in a hash-chained audit log. Blast-radius containment for the day an agent goes wrong, because one will.

github.com/WeLikeCode/mintkey

mint-mory: memory your agents can actually trust

Typed, self-consolidating memory for LLM agents in a single SQLite file. Eight memory categories, a concept graph, contradiction resolution ("dreaming"), MCP-native. Benchmarked against purpose-built alternatives.

github.com/WeLikeCode/mint-mory

Why this matters for your review

When I flag a credential or memory risk in your architecture, it's not theory; I've hit the same wall building the reference implementation. That's the difference between advice and experience.