Getting started
Core concepts
Short definitions of the terms used throughout the documentation.
| Term | Meaning |
|---|---|
| Agent | The software that calls tools. Each agent has an identity (agentId) and its own policy. |
| Session | The context formed by an agent's consecutive calls. Trajectory analysis is per session. |
| Tool call | An agent's request to run a tool with given arguments: { tool, args }. |
| Policy | A YAML file. States which tools are allowed, which arguments are constrained and what requires approval. |
| Decision | The outcome of every call: { decision, reason, ruleId }. |
| Request phase | Before the tool runs: policy + egress scan + intent analysis. |
| Response phase | After the tool runs: prompt injection scan of the returned result. |
| Trajectory | The ordered call history of a session; the basis for loop and slow-exfiltration detection. |
| Upstream | The real tool server behind AvertGate (an MCP server or HTTP API). |
| Fail-closed | If an inspector errors or times out, the call is blocked or sent for approval — never allowed. |
| Tenant | A customer's private, isolated workspace: its dashboard, agents, keys and policies. |