At a glance
- Answers from your platform data, with citations.
- Proposes changes; nothing runs until you confirm.
- Never more access than the person asking.
What it can do
- Answer from your data. It reads your deployments, their metrics and configuration, your models, usage, trial credit, API keys and activity. Each answer shows which records it relied on.
- Explain and recommend. It compares before-and-after metrics, explains a deployment's effective configuration, and suggests next steps.
- Propose changes (Preview). When your Tenant allows it, the FDE can propose a short list of deployment actions: deploy a model, scale a deployment, tune its settings, or retire it. Reversible actions can be rolled back.
How it stays inside your permissions
- Your access, never more. The FDE works through the same public API you use, with no more permission than the person asking. It can't reach another Tenant or outside your Organization, whatever a prompt or a document tells it.
- Read-only unless you say otherwise. Out of the box the FDE only answers. A Tenant admin chooses whether it may propose changes, and can switch it off entirely.
- Nothing runs until you confirm. A proposed change appears as a plan. The platform builds the plan from your data; the model doesn't write it. Each step needs your confirmation in your session, a confirmation can be used only once, and plans expire if left unconfirmed. Irreversible actions, such as retiring a deployment, get their own confirmation.
- Checked twice. Every step is checked against policy again just before it runs, with platform limits on what it may change.
- Recorded. Questions, proposals, confirmations and actions appear in your activity trail, showing the FDE acting on behalf of the person who confirmed.
Guardrails
- Checks at every stage. The FDE screens what you ask, what its tools return, any action before it runs, and what it says before you see it. If a required check can't run, the turn stops.
- Instructions in data are treated as data. Text in a tool result or a document that tries to give the FDE instructions is ignored as an instruction. Attempts to exfiltrate secrets or reach other Tenants are blocked or redacted.
- Bounded work. Each turn has hard limits on steps, tool calls, time and tokens, and each Tenant has limits on concurrent and daily use.
- No open internet. The FDE can reach only the rackAI API and gateway.
What it can't do
- Change anything without your confirmation, or anything outside its short list of deployment actions.
- See or touch another Tenant, or anything your role doesn't allow.
- Browse the web, run code, or call outside services.
- Act on its own schedule.
Your conversations
Conversations are saved so you can return to them. Only you can see yours unless you share one, and you can delete them.