Skip to content

Glossary

The product's terms, and the distinctions behind them. Several pairs are easy to conflate and OpsPilot treats them as genuinely different: air-gapped against private or VPN-only, a connection against a session, AI Providers against AI Assistants. Where a term has a page of its own, it is linked.

Air-gapped
No network route to the internet at all. OpsPilot is only fully air-gapped when the model runs locally too — see Offline with Ollama. With a cloud provider the workstation still reaches the internet, which is private or VPN-only, not air-gapped.
AI Assistant
An external application you already use — Claude Desktop, ChatGPT Desktop, ChatGPT Web, VS Code Copilot Chat — connected to OpsPilot over MCP. It uses the subscription you already pay for rather than an API key. See AI Assistants. Distinct from an AI Provider.
AI badge
The per-connection indicator showing whether the assistant may see that connection at all. It is not rendered for external launchers, file browsers, Telnet or RSH, because those sessions cannot use AI regardless.
AI Provider
A model backend you configure inside OpsPilot with an API key or a sign-in, driving OpsPilot's own AI panel. Ten ship. See AI Providers and the provider matrix. Distinct from an AI Assistant.
Approval gate
The human decision between a proposal and execution. It is the point every AI path converges on, whichever model or assistant produced the proposal. See Approvals & auto-run.
Auto-run
The optional setting that lets read-only commands run without an approval click. It ships off, and it is a single application-wide switch rather than part of a profile.
Command Safety profile
A named policy holding dangerous patterns and the dangerous-confirmation setting. Resolves connection → group → Default, and replaces rather than extends. See Command Safety profiles.
Connection
A saved definition of something to connect to — type, host, credentials, environment, group and AI scope. Creating one does not open anything. The open thing is a session. See Add a connection.
Dangerous pattern
A plain-text, case-insensitive substring that forces a proposed command to the High risk tier. Deliberately not a regular expression. A match can only add the dangerous classification, never remove one the model already set. See Dangerous patterns.
Data Handling profile
A named redaction policy: which of the ten built-in categories are on, plus your own regular expressions. Resolves connection → group → Default. See Data Handling profiles.
Detached session
A session moved into its own window, so a terminal can sit on a second monitor while the AI panel stays on the first. Some main-window features do not follow it. See Sessions & tabs.
Environment
A colour-coded tag such as production or staging, used for visual separation and as a unit for scoping policy. production and staging are seeded on a new install. See Groups & environments.
Execution boundary
The architectural property that the AI has no execution capability — its only output channel is a proposal. Not a setting; the product reports it as Always enforced. See Execution boundary & risk tiers.
Group
A folder of connections that can carry policy. Assigning a Command Safety or Data Handling profile to a group applies it to everything inside, which makes groups the efficient place to set policy.
Hypervisor console
A VNC connection attached to a virtual machine's console through its hypervisor — KVM/libvirt over SSH and virsh, or OpenStack Nova over Keystone and Nova — rather than to a VNC server inside the guest. What you need when a machine will not boot or has no SSH. See Hypervisor consoles.
MCP
Model Context Protocol, the standard external assistants use to reach tools. OpsPilot exposes itself as an MCP server bound to 127.0.0.1 only. See MCP tools.
Private / VPN-only
The deployment most regulated teams actually need: the targets are isolated, and the OpsPilot workstation reaches both them and a cloud AI. Not air-gapped — keep the two apart when describing what a deployment achieves.
Profile precedence
The order in which a profile is resolved for a session: the connection's own, then its group's, then Default. Because profiles replace rather than extend, a more specific profile can be less strict than Default.
Proposal
A command the AI suggests. It has not run and will not run until a person approves it. Never describe a proposal as something the AI did.
Redaction
The local removal of secrets from text before it reaches any AI, replacing each match with a typed placeholder. Every route to a model passes through it first. See What gets sent.
Redaction badge
The 🔒 indicator, with a count, on any turn where something was scrubbed; hovering it names the categories. Hiding the badge hides the indicator only — redaction still happens.
Risk tier
One of three classifications for a proposed command: Read-only, Low risk or High risk. Determined by the model's own assessment combined with your pattern list, where the list can only make the tier stricter.
Secure tunnel
The outbound connection that lets browser and mobile ChatGPT reach OpsPilot. It forwards only to a validated loopback address with a token, and opens no public listener. See Remote & mobile operation.
Session
An open connection — a terminal, an embedded desktop or a file browser, in a tab. Sessions are what the AI can see, what proposals target and what profiles resolve against. Distinct from a connection.

See also