An open-source experiment in agent controls
Before an agent acts,
what should it be allowed to do?
Explore how a proposed action meets a specific set of controls. Change the scenario. Inspect the decision, its binding, and what the outcome actually proves.
Recorded reference examples. These outputs come from the real Python evaluator. This static page does not evaluate new policies or execute actions. An “allow” result here is not production permission.
-Choose a proposed action
Selected immutable example pack
ControlSpec reference result
Decision and route
Recheck the exact action
Assess the outcome
Before an
agent acts.
A permitted connection is one part of the decision. Inspect the exact action, its approval, and the evidence it leaves behind.
This page replays recordings. Scripted inputs exercise the real ControlSpec evaluator through an authenticated gRPC fixture and a synthetic service. No live agent or real transactions run in this browser. Test operator approvals are simulated.
One request. Three control paths.
Inspect the exact scripted request
Inspect the recorded data and provenance
Download all seven observationsBuilding an agent? Try one of your controls against the Python reference. We want integration attempts, uncomfortable edge cases, and feedback on the contract.
Share an integration idea ↗What was actually verified
- OpenShell pin
- Runtime run
- Runtime detail
- Verification
- Source revision
- Recorded
Read the boundaries, too.
Inspect the selected scenario’s raw trace
Requests, decisions, target state, and available evidence are displayed as recorded. The full download includes provenance and every scenario.
Download all recorded traces ↓This is a bounded experimental complement to NVIDIA OpenShell, with no implied endorsement. OpenShell itself supports body-aware controls; the value of this example is its explicit action binding, approval path, and evidence. Inspect the configuration before comparing outcomes.
Challenge the boundary ↗