Where should an agent permission check live in your stack?

I’m Sean, the builder behind NextEleven and DeadBolt. I’d like to hear from developers about the hardest part of adding execution controls to an agent they already use.

DeadBolt puts a permission check at the tool boundary. A trusted executor checks the gate and starts the protected action only after a fresh explicit allow. You can build it in with Rust, or bolt it on through a local HTTP sidecar or trusted stdio MCP proxy. It is free and open source under Apache-2.0.

The boundary matters: it protects the routes you connect to it. It is not a whole-agent sandbox and cannot cancel an action already running.

The first evaluation uses harmless temporary files and needs no AI account. The captioned walkthrough shows it inside Harness.

Source and evaluation guides: https://github.com/seanebones-lang/deadbolt

Captioned overview:

Where would this fit in your stack, and what would make it easier to try? A middleware hook, an HTTP service, an MCP proxy, or something else? If you try the evaluation, I’d value hearing which step was unclear.