NullRun - Runtime Authorization for AI Agents - Before They Execute
by•
NullRun authorizes AI agent actions at runtime.
Each action is evaluated as allowed, requires human approval, or blocked, with evidence recorded for every decision.
Python SDK, hosted control plane, approvals, RBAC, traces, audit trail, and tenant isolation are live. Self-serve and deploys in minutes.
Replies
Hey Product Hunt 👋 I'm Anatolii, the solo founder behind NullRun.
I didn't start with a big "AI governance" thesis.
I started with a much more annoying engineering question: if an agent is allowed to use a tool, who actually gets the final say when it's time to execute?
NullRun sits right at that execution boundary - a control point that checks an agent's tool call before it actually runs, instead of trusting the agent to decide for itself. The first version was literally a Python decorator around one function. I kept pushing the same idea closer to the real point of execution until it became a product.
What interests me now isn't whether agents can make mistakes - we already know they can.
It's what happens when an agent has enough access for a mistake to actually matter.
I'm also building this solo, so I'd genuinely rather learn from people here than just talk about what I built.
If you're working with agents: what's one action you'd trust an agent to request, but still wouldn't trust it to execute on its own? Real cases welcome, especially the messy ones.