| Agent | What Xenovia governs. It has an owner, a purpose, a primary provider and up to three fallbacks, policies, guardrails, optional sub-agents and its own API keys. All governance is scoped to an agent. See Create an agent |
| Agent API key | An xe_ key for one agent’s endpoint and SDK traffic. See Users and keys |
| Organization API key | An ak_ key for the management API |
| Hooks key | An agent key that works only on the coding-agent hook endpoints. See Coding agents |
| Purpose | What an agent is for, in plain language. Every model request is checked against it |
| Capability | A named thing an agent does. When an agent has capabilities, a request must match one |
| Provider | Holds the upstream credential and model. An agent sends its requests to its primary provider and can fail over to its fallbacks; credentials are resolved server-side. See Providers |
| Fallback | A backup provider, tried in order when the primary is down, overloaded or rate-limited. See Fallbacks |
| Policy | A Rego rule evaluated by OPA. Request-stage decisions are allow, block, redact or escalate; response-stage decisions are allow, block_response, redact_response or escalate. See Policies |
| Neon | Xenovia’s decision model. It checks every stage for its trained risks, your guardrails and the agent’s purpose. See Neon |
| Guardrail | A named check you write for an agent, in plain language, evaluated by Neon. The scope guardrail limits what requests may ask the agent to do |
| Protection | What one of Neon’s trained risks does when it fires: Enforce, Monitor or Off |
| Monitor / Enforce | The two modes of every control. Monitor records what it would do and changes nothing; Enforce acts |
| Stage | Where a check runs: model request, model response, tool call, tool result, memory write or context summary |
| Hosted tool | A tool the model provider runs, such as web search or a remote MCP server. See Hosted tools |
| Session | A group of related turns, resolved per request or set with X-Xenovia-Session-Id. The Sessions page lists these |
| Turn | One request/response cycle through an agent, or one SDK check. A turn carries the policy decision, tokens and latency. Some screens call a turn a “trace” |
| Run / step | A unit of work inside a turn: the model call, a policy check, a runtime-executed tool |
| Workflow | Turns across several agents grouped by a shared workflow ID, for multi-agent runs |
| Sub-agent | A step inside an agent with its own keys and extra guardrails (inline), or another agent this one may call (linked). See Sub-agents |
| Escalation | A step queued on its session for human review and resolved with a note. A policy escalation refuses the call like a block; a guardrail escalation lets the call proceed |
| Refusal | The 200 reply a blocked call gets, in the format of the API you called. See What a block looks like to your app |
| Discovery | The list of agents and AI apps your organization uses. See Discovery |
| Finding | One thing wrong with a Discovery entry |
| Shadow AI | An app people use that nobody reviewed and no connected integration can govern |
| Hybrid deployment | The Xenovia runtime running in your own cluster. See Hybrid |
| Nova | The console’s helper. It looks things up and suggests changes; a person applies them. See Nova |