What goes wrong
Putting something in front of OpenAI, Anthropic, Gemini, or Azure is the right place to measure wait time, errors, and volume. The usual failure is retention: full chats land in a log stack “just for debugging,” and stay there longer than anyone intended.
The other choice is to skip the middle layer and stay blind — no shared picture of what tools are actually calling.
Why it matters
You are asked for both a paper trail and privacy. A cloud logger that holds prompts fails the second. An editor-only dashboard fails the first when traffic spans tools and providers.
What good looks like
A local path that records operational facts (which service, which model, status, rough size, wait time) while chat contents stay in memory for the request — plus a way to verify that prompts are not written to disk.
How Aura Wire helps
Aura Wire inspects traffic only to the AI services you allow. Everything else passes through untouched. What it keeps is sanitized events — not message text.
- Live traffic and request detail from metadata only — shapes and sizes, not bodies
- You control how long those records stay, with cleanup on your machine
- A built-in privacy check that fails if prompts land on disk
Try Aura Wire free
AI chat for VS Code with a local pipeline behind it — live spend, secrets stopped, nothing stored.