A cage on the agent, or a check you can take with you?

Both, but they are not the same thing.

A sandbox cages what an agent can touch: files, the network, money. That matters. NVIDIA OpenShell cages what an agent can touch. TrustShell checks what it says.

A cage stays where it was built. A check should go where you go. TrustShell is a portable trust harness: the same check on this page, in ChatGPT, Claude, Gemini, Grok and DeepSeek through the browser extension, in your terminal, and in any agent that speaks MCP. It works whatever model wrote the answer, so you are not locked to one vendor.

You set it. Choose which chat sites it checks, and whether it checks every reply or only when you click. Your record of stamps stays in your browser.

It bends instead of breaking. If a checker is busy or down, a backup from the list in our privacy policy takes its turn, and the two answers that decide a stamp always come from two different model families. If no pair can answer, the stamp says Not checked. It never passes by default.

Glass box means you see who checked the sentence and what they said. It does not mean the sentence stayed on your machine: to be checked, it is sent to the checkers named in our privacy policy, and it is not stored.

We built it because the alternatives are waiting: for a few big labs to grade their own work, or for regulators to catch up. A check you carry is yours today.

People, helping agents, to help people, help people.

How the two-family check works, in depth: disjoint-family verification. The whole arc — earn trust, learn skills, unlock autonomy inside a signed ceiling — is on the trust harness.