# Glossary

**Admin** — someone who uses the console. Manages policy, people, providers and billing.
Consumes a seat.

**Agent** — an AI assistant that can act, not only answer, by calling tools.

**Blocked** — a request or tool call your policy stopped. Not an error.

**Developer** — someone whose AI tools are governed. Uses the Endpoint Suite, never the
console. Consumes a seat.

**Device** — one enrolled machine, identified by a key pair in its secure storage.

**Drift** — a governed tool that no longer points at the Gateway, because its configuration
was changed outside the app. Surfaces as **Non-compliant**.

**Endpoint Suite** — the desktop app that enrolls a machine and points its AI tools at the
Gateway.

**Evidence pack** — the point-in-time report from **Compliance**.

**Gateway** — the proxy every governed AI request passes through.

**Lethal trifecta** — private data access, untrusted content, and an exfiltration channel
present together. See [Lethal-trifecta enforcement](/books/policy-and-risk-detection/page/lethal-trifecta-enforcement).

**Managed API key mode** — requests billed to your organization's provider key.

**MCP** — Model Context Protocol, how an assistant calls tools.

**Posture finding** — something about an AI tool's *own* autonomy settings on a device.
Report-only.

**Prompt capture** — the opt-in setting that stores conversation content.

**Routed endpoint** — a remote MCP server reached through the Gateway, so policy applies.

**Rug pull** — an MCP tool whose description or schema changes after you approved it.

**Seat** — one person, admin or developer. Shared pool.

**Shim** — the small process between an AI tool and a local MCP server that asks the
Gateway for a decision.

**Slopsquatting** — registering package names that AI models commonly hallucinate.

**Subscription mode** — requests billed to the developer's own AI plan.

**Ungoverned agent** — an autonomous AI system on a device whose traffic does not pass
through the Gateway, so it can be detected but not routed, enforced or audited. See
[Ungoverned AI agents](/books/devices-and-logs/page/ungoverned-ai-agents).