FiloVenturini

Previous Show HN: https://news.ycombinator.com/item?id=47136585. That version had the vault abstraction but access control lived entirely in application code — which is exactly the problem I was trying to solve.

This version adds server-side enforcement: access rules are declared via CLI and checked on every request, independent of what the agent implementation does. The boundary doesn't disappear if you write a bug.

Also added MCP server support, so Claude Desktop can use it with no integration code.

Have you considered adding any kind of agent coordination layer, e.g. letting one “orchestrator” agent spawn and direct sub-agents on specific subtasks, rather than having the developer manually assign each task? Or is the explicit human-in-the-loop assignment a deliberate design choice to keep control and avoid runaway costs?

I’m also wondering whether there is actually a need for memory to become a controllable infrastructure layer that agents can integrate with, instead of everyone building custom memory management logic each time.

Do you think this is something that will become necessary in production agent systems, or is memory still mostly treated as an implementation detail today?