You sit between tools and the model already. Would an authority check belong in the same seat?

[javmann100]

I maintain mnki, an open-source trust layer that decides, before a tool runs, whether the agent was allowed to run it, and records why. Headroom sits in the adjacent seat: a proxy or wrap in front of fifteen agent tools that compresses what comes back from tools before the model sees it.

That seat is the one I keep coming back to. headroom wrap claude|codex|cursor|… gets you in front of the tool loop in one command with no code changes; the same position is where "may this agent make this call at all" and "what did it do" would live. I am not proposing Headroom should do that; I am asking whether the people who install a wrap for tokens ever ask you for control, or whether cost is the only reason anyone accepts a proxy.

Twenty minutes, or a reply here, would help me decide where the check belongs. Integration paths, including an observe mode that changes nothing: https://mnki.com/docs/integrations. Thank you for Headroom; keeping the FATAL line while dropping ninety percent of a log dump is the demo that sells the idea.