How It Works

You don't need this page to use your agent — but if you're the person your organisation asks "where does the data actually live?", this is for you.

The shape of the system

Three parts:

1. The device — an appliance at your facility running the agent and its supporting services. All of the agent's data (conversations, memory, files, wiki, settings) lives here.
2. Performance Hub — the management plane. Pairing, access control, budgets, backups configuration, and updates are all decided here and delivered to the device.
3. Cloud services the device consumes — AI models (via your facility's AI Models & Access account), encrypted backup storage in your chosen regions, and push notification delivery.

The guiding principle: your data stays on your device; the cloud provides brains, backup, and reach.

Outbound-only networking

The device makes only outbound connections — ordinary HTTPS and secure tunnels dialled out from your network. It never requires an open inbound port, port-forwarding, or firewall changes. If your network can browse the web, the device can do its job.

This is also how Performance Hub reaches the device: the device dials out and holds open a secure tunnel to Performance Hub's relay infrastructure, and management traffic and remote access flow back down that tunnel.

How your browser reaches the agent

When you open the AI Agent module, what you see is the agent's own web interface, served from the device. Your browser connects over the best available path, and the header shows which one:

The module tries them in that order automatically; you never choose. All three paths are encrypted end to end, and all three carry your Performance Hub identity, so the agent always knows who it's talking to and access control is enforced identically on every path.

Sessions to the device use short-lived signed tokens minted by Performance Hub when you open the module — there are no standing passwords to the device, and access revoked in Performance Hub takes effect when the current short-lived session expires or the page reloads.

How the agent gets AI models

The agent's intelligence comes from AI models delivered through your facility's AI Models & Access account. On the device, a small proxy service sits between the agent and the cloud:

• The agent is pre-configured with an auto model selection — Performance Hub routes each request to an appropriate model, and improvements roll out without you re-configuring anything.
• The device authenticates with short-lived signed tokens derived from a per-device secret issued at pairing. There is no long-lived API key stored on the device — nothing that would remain useful if the hardware were stolen.
• Every request is metered against the agent's own spend key, which is how Budgets & limits are enforced — centrally, on the Performance Hub side, not on trust.

How the agent reaches Performance Hub data

The agent's access to your business data works the same way, over MCP (Model Context Protocol):

• A proxy on the device forwards the agent's MCP calls to Performance Hub, authenticated with the same short-lived device tokens.
• On the Performance Hub side, the device acts as a service account whose scope and capabilities are exactly what you granted under Connecting to Performance Hub. Every call is authorised centrally against that grant.
• Changing or revoking the grant takes effect within moments — the device holds no independent credentials to your data.

How backups flow

A backup service on the device takes encrypted snapshots hourly and uploads them directly to cloud storage in your chosen regions — encrypted on the device before upload, with the encryption key held by Performance Hub for your facility (never stored in the backup itself). Backups are incremental and deduplicated, so hourly cadence stays lightweight. Details on Backups.

How updates work

The device runs a managed software stack, updated as a whole:

• New releases are staged to the device automatically, but held until you click Update now or schedule a time — the device never restarts itself on you unannounced.
• Updates replace the software, not the data: the agent's state lives in a persistent data area that carries across every update.
• If you never touch the banner, the device simply keeps running its current version. That said, we recommend keeping devices on the latest version — every update is rigorously tested before it is rolled out, and staying current means you always have the newest capabilities and fixes.

If the internet goes out

• You at the facility: the Direct (local network) path keeps working — you can still open and use the agent over your LAN, though it can't call AI models (so it can't think about new requests) until connectivity returns.
• Backups and notifications queue up and resume when the connection is back.
• Performance Hub management (drawer, settings, updates) needs the device online and reconnects automatically.

If the hardware dies

Backups are the answer: cloud restore points remain browsable and downloadable even with the device dead, and Replace this device stands up a new unit from the old one's latest backup — typically losing at most an hour of state. See Restore & recovery.

Related pages

• The hardware
• Security & privacy — the same ground, organised as a security review
• Connecting to Performance Hub
• Backups