Glossary
The words Reins uses, each in a sentence or two, with a link to the page that explains it.
Activity. The tab of the panel that lists the last 50 commands: time, agent, method or action name, outcome, and what let an action run. Never arguments, results or page text. See Permission levels.
Adopt. To drive an Electron app you did not write: reins-electron loads the driver into its window through the app's debugging connection and carries its frames. An attached agent can then press anything in that window. See Adopt an Electron app.
App origin. The origin of a page on an app's own scheme, written app://name. * in an origin list does not cover it, nor file:// or null: each has to be named. See Handshake, authentication and sealed sessions.
Action. Something the app can do, declared with a name, a schema for its arguments and a risk tier, and called by an agent with act. See Writing good actions and risk tiers.
Agent. An AI harness, or an assistant running in another tab or in the same page. One of the two roles a connection can have. See Roles and trust.
Attach. To ask for a grant on a surface, with surface.attach. The page decides, and under the guarded profile the person is asked. See Grants and scopes.
Audit log. The hub's append-only, hash-chained record of connections, grants and requests. See The audit log.
Always allow. The third button of a consent card for an action. It sets the person's rule Allow for that action and agent, in two presses. See Consent.
Broker. The part of the hub that authenticates peers, issues grants, enforces policy and routes requests. It is transport-agnostic, so a page can host it too.
Canonical JSON. JSON without whitespace, with sorted keys and numbers written one way, so that two parties hash the same bytes. See Manifest and prompt card.
Chat. The page's one chat room, in the Chat tab of the panel: the person at the page and every attached agent whose grant holds the chat scope. A message is a request, never an approval. See Chat.
Chat relay. Whether agents in the chat hear each other: the policy field chatRelay, off, person (once the person switches it on) or on. The stricter of the page's and the hub's applies.
Confirm step. The second step the panel shows before anything that lets an agent do more: it names the agent and what stops asking, and its button waits 500 ms. See Permission levels.
Consent prompt. The card the driver shows when policy answers confirm, for an attach or a sensitive action, with Deny, Allow and, for an action, Always allow. See Consent.
Credential. What an agent authenticates with: a key id (kid), a key, and optional caps on scopes, origins and sessions. See Handshake, authentication and sealed sessions.
Cursor. The agent's pointer, drawn by the driver so the person can follow. Visual only. See Cursor and the pointer stream.
Decision. What policy answers: allow, confirm or deny.
Delta. What changed in the UI tree since an agent last looked: a list of set and del operations. See Observation, handles and deltas.
Driver. The in-page runtime of the reference implementation: the UI tree, command execution, the cursor, the consent prompt and the indicator. See Make an app drivable.
Exposure. Which elements the UI tree shows: auto, annotated or none.
Grant. An agent's permission on one surface: a set of scopes and an expiry. See Grants and scopes.
Handle. The id of an element in the UI tree, such as e12, stable for the element's lifetime.
Hub. The broker both sides connect to, and the process that binds it to a WebSocket on loopback.
Indicator. The pill the driver shows while an agent is attached, naming it, with Pause and Stop. Its text opens the panel, and it shows a level other than Ask and the person's Allow rules. See Human control.
In-tab shape. Reins with no hub process: the page hosts the broker, and the assistant and the driver talk over in-memory links. See The in-tab shape.
Kit. The single entry point of the package for browser projects, @offensiveadhd/reins, and its bundle, reins.js.
Launcher. reins.mjs in the Reins folder: a file that runs the shell tools of the installation that wrote it. See Shell tools with nothing registered.
Level. How much the person is asked about one agent's actions, set per grant in the panel: strict, ask (where a grant starts, unless the person set that agent to Strict earlier on the page), auto or bypass. See Permission levels.
Link. A bidirectional, ordered, reliable message pipe. WebSocket, MessagePort and the in-memory pair all present this shape. See Transports and framing.
Manifest. A surface's description of itself: its app, actions, views, state, exposure and profile. See Manifest and prompt card.
Notice. A line the MCP adapter and the shell tools put before a tool's own answer, starting [reins]: what the person set, how many chat messages wait. Never page text. See Use it from Claude Code and other MCP harnesses.
Overlay. The indicator, the panel and the prompts: the driver's own controls, in a closed shadow root in the browser's top layer, taking only the person's own input.
Panel. What the indicator's text opens, for the person alone: the tabs Permissions, Activity and Chat. See Permission levels.
Pointer stream. A binary channel for live cursor trajectories, with credit-based flow control. See Cursor and the pointer stream.
Policy. A profile plus overrides: the rules that turn a request into allow, confirm or deny. The hub has one and each surface has one. See Policy profiles.
Private field. An element an agent must never read and, by default, never write: passwords, card numbers, one-time codes, and anything marked data-reins="private". See Privacy annotations.
Profile. One of three preset policies: open, guarded or kiosk.
Prompt card. The short text, rendered from the manifest, that teaches a model to drive an app. See Manifest and prompt card.
Reins folder. ~/.reins, or the folder REINS_HOME names: where a hub announces itself and leaves the launcher.
Risk tier. An app's statement about an action: safe, sensitive or dangerous.
Rule. The person's decision for one action of one agent, set in the panel: Allow, Ask or Block (allow, confirm, deny). It beats the level and covers any arguments. Allow ends with the grant; Ask and Block stay with the agent for as long as the page lives. See Permission levels.
Scope. A kind of thing an agent may do: observe, act, ui, cursor, show, nav, chat.
Scope switch. A switch in the panel that turns one scope of a grant off, and back on, for one agent. A request whose scope is off fails with E_SCOPE. See Grants and scopes.
Sealed session. A session whose every message carries a MAC under keys derived from the shared secret and the handshake. See Handshake, authentication and sealed sessions.
Semantic id. The name an app gives an element with data-reins-id, addressed as #name.
Session. A connection's identity at the hub (a1, s1): a label, never a credential. Also the background process the shell tools keep, which holds a connection and its grant between commands.
State. A named piece of app state an agent may read, with a getter in the app.
Surface. A web page that embeds the driver. The other role a connection can have.
Surface key. A secret shared by the hub and a page, by app id, with which each proves itself to the other. See Surface keys.
Target. How a request names an element: a handle or a semantic id, never a selector.
Token file. The file that holds an agent's secret, by default .reins/token where the hub was started.
Unverified hub. What a page calls a hub whose proof it cannot check, because the page holds no key. Such a page never accepts an attach silently.
View. A named renderer the app offers for showing data, fed by show.view.