Observability and governance for 30 runtimes: which of the "controls" do users actually reach for?

[javmann100]

I build mnki, an open-source trust layer that decides, before a tool runs, whether the agent was allowed to run it, and records why. It is adjacent to Clawmetry rather than competing: you show what each runtime did, turn by turn, reading their logs read-only; mnki is the check that sits in front of the call.

Your docs are unusually honest about which controls are real per runtime, and that is exactly what I want to ask about. When someone watches a session replay and sees a tool call they did not expect, what do they do next: block that tool in the runtime, add an approval, or shrug because none of the runtimes give them a real lever? And is the approvals story something users ask you to extend, or a nice-to-have?

Twenty minutes, or a reply here, would help me decide whether that lever belongs in a layer like mine or in the runtimes. Integration paths: https://mnki.com/docs/integrations. Thank you for Clawmetry; the per-runtime map of what cannot be seen is the most useful page in the repository.