Joint Reference Architecture

Wand AI gives an autonomous agent an identity, a manager, goals and an audit trail. Zeroport gives that agent a body it cannot escape from. Every place where a Wand agent leaves the platform — into a sandbox’s egress path, into a protected internal estate, or into a model — the traffic terminates on a Zeroport Fantom Core: a sealed, per-session appliance with a physical break in the wire.

The joint architecture is two layers with a single, repeated pattern.

Wand (application layer) decides what the agent is authorized to do, and routes every outbound action through one of three gateways.

Zeroport (hardware layer) decides whether it physically can, and enforces that from silicon the agent has no address for.

No IP, TCP, UDP or proprietary packet crosses the boundary. What crosses is human-interface input inward — keyboard and mouse, carried as analog signals — and a display-only pixel stream outward.

The path is physical. A Fantom Core is a circuit board roughly one-third the size of a credit card, carrying two SoCs separated by a full physical air gap. The gap terminates no network session and carries no packet. There is nothing to enumerate, nothing to pivot into, and no integration to secure.

The policy is out of band. Moativ, Zeroport’s on-device AI, runs on NVIDIA Jetson Thor NPUs inside the appliance — one internal, one external. It sees every keystroke going in and every pixel coming out on a single timeline, and distinguishes a command merely typed from one executed from one already showing results. It can alert, block, or physically sever a session via a kill command that is completely out of band. A software guardrail runs on the same computational plane as the model it polices; Moativ does not.

Policy is written in plain language, compiled into staged detectors, and dry-run against the program’s own session history before it is armed. Autonomy is set per group of users or agents on a five-step dial: observe, alert, investigate, terminate, self-tune. Every autonomous action lands in the audit log with the evidence attached — what was on screen, what was typed, what was done, and why.

Nothing leaves the rack. All inspection is on-premises. Zeroport sees no session content, and neither does any other external party.

Zeroport is a boundary, not a control plane. Identity, authorization, objectives, the agent’s place in the org graph, the tool and application catalog, the classifier and the deny / approve / capture-approval flow, and the audit system of record all stay in Wand. Zeroport consumes the authorization decision and makes it physically true.

Rule binding is granular enough to follow Wand’s model: per circuit board, per user or agent credential, or round-robin across the available pool. Identity comes from the program’s existing IdP, MFA and AD.

Scale: 200 Fantom Cores per 1U appliance; up to 10 appliances daisy-chained for roughly 2,000 concurrent sessions. Capacity grows in Cores, then in appliances — there is no shared tunnel to bottleneck or harden.

Latency: dedicated hardware pathways target sub-50 ms, which is what makes the channel usable for live control work rather than only for after-the-fact review.

Commercial model: per concurrent session, hardware-as-a-service, not per named licence.

Human operators and agents ride the same rails under the same policy engine.