SentnelOpsGlossary
Every term you'll meet in SentnelOps, defined in plain English. Bookmark this — the rest of the docs assume nothing more than what's on this page.
Terms appear roughly in the order you'll encounter them: first the things being governed, then the credentials, then the governance machinery itself.
| Term | Plain-English meaning |
|---|---|
| AI agent | Any piece of software that uses AI to act on your behalf — a coding assistant, a chatbot, an automation. In SentnelOps, each one gets a registered identity, like an employee record. |
| MCP | Model Context Protocol — the standard language AI agents use to ask tools to do things. Think of it as the common plug shape that lets any AI connect to any tool. |
| MCP server | A service that offers tools over MCP — the thing that actually reads files, queries databases, or sends messages when the AI asks. |
| Tool | One specific action an MCP server offers, like “read a file” or “delete a record”. Policies are written about tools. |
| Agent token | The credential an agent shows on every call — its ID badge. It only works at the proxy, is shown once when issued, and can be cancelled (“rotated”) in seconds. |
| Org API key | The master credential for your organisation's account (starts with snops_key_). It's for scripts and setup, not for agents — keep it in a password manager or secret manager. |
| Registry | The directory of every AI agent in your organisation: who owns it, what it may touch, and whether it's currently allowed to run. The employee directory, but for AI. |
| Policy | A written rule saying which tools an agent may use. Like a building access list: this badge opens these doors and no others. Your organisation edits its own rules on the dashboard's Policy page. |
| Blast radius | A score from 0.0 to 1.0 for how much damage a tool could do — reading a file is near 0, deleting a database is near 1. High-blast actions can be made to require human approval. |
| Observe mode | The default. Every call goes through untouched, but everything is recorded — including what the rules would have stopped. The security desk taking notes without stopping anyone. |
| Enforce mode | The same rules, now with teeth: calls the policy denies never reach your tools. You turn it on per environment, deliberately, after watching observe mode. |
| Would-block | A flag on each logged call meaning “if enforcement were on, this would have been stopped”. It's a free, exact preview of what enforce mode will do — using your real traffic. |
| Approval | A human yes/no on a risky action or a production agent. The AI waits; a person decides; the decision is recorded with their name. |
| Ghost agent | An agent that's running but was never registered — software calling your tools with no badge on file. The dashboard's Ghost agents page hunts for these. |
| Audit trail | The permanent record of every call and every decision. Append-only: entries can be added but never edited or deleted — a logbook in permanent ink. |
| Lifecycle | The stages of an agent's existence: draft → approved → active ⇄ suspended → decommissioned. Like an employee going from offer letter to onboarded to on-leave to departed. |
| Roles | What each human teammate may do, in increasing order of power: viewer (look only), developer (register agents), security (lifecycle and approval decisions), admin (manage people). |
| Proxy / gateway | The checkpoint itself: the address your AI's calls pass through on the way to your tools. It identifies the caller, consults the rules, writes the logbook, and forwards the call unchanged. |