AI glossary

Plain-language definitions of the terms behind on-device AI agents.

Local AI agent

An AI assistant that runs on your own device and performs real-world tasks — operating websites, signing in, downloading files, and sending messages — without uploading your data to a remote server.

On-device AI

Artificial intelligence that processes information on your own computer or phone rather than on a company's cloud servers, keeping your data private and local.

AI agent

Software that takes actions to complete a goal, not just answer questions. Unlike a chatbot, an AI agent operates apps and websites to finish a task.

Autonomous task execution

The ability of an AI assistant to carry out a multi-step task from start to finish, such as logging in, downloading a document, and sending it, with minimal manual input.

Human-in-the-loop

A safety design where the assistant pauses for a person's approval before any critical or irreversible action. BackBrain uses this before submitting filings or sending files.

Cloud AI assistant

An AI assistant that runs on remote servers and requires your data to be sent to it over the internet in order to help you.

Visual grounding

A technique where an AI identifies on-screen elements — buttons, fields, menus — by perceiving the screen, allowing it to operate websites that have no special integration.

Autonomous agent / AI operator

Software that perceives, decides, and acts in a loop to complete a goal — as opposed to a chatbot, which only produces text.

Local-first

An architecture where the application, your data, and its learning live on your own device rather than a vendor's cloud.

Autonomy tiers

BackBrain's three-level action classification — routine read-only steps run automatically, sensitive steps take a one-tap approval, and irreversible or financial steps always require explicit confirmation.

Safety floor (deny-floor)

The deterministic rule layer, separate from the AI model, that blocks forbidden actions (like unapproved deletion) and that the model cannot override or talk its way around.

Audit ledger

BackBrain's tamper-evident record of agent actions, where each entry is cryptographically chained to the previous one so alterations are detectable.

Run transcript / History

The complete, searchable, secrets-redacted record of a single run — everything said and done, kept on your device.

Encrypted vault

On-device credential storage protected by Windows' built-in encryption (DPAPI); secrets are supplied just-in-time and never written to logs.

Session-only credentials

An option to hold a password in memory for one task only — used, then cleared, never stored.

Redaction

The automatic removal of sensitive data from anything sent to a cloud AI model; credential screens are never transmitted at all.

Local-only mode

A privacy setting where all AI reasoning runs on a local model and cloud egress is blocked — with no silent fallback to a cloud provider.

OTP relay

The automatic hand-off of a one-time passcode from your paired Android phone (or your email) into the desktop task that's waiting for it.

Compiled procedure (instant replay)

A repeatedly verified task distilled into a step sequence that replays with zero AI calls — and hands back to live reasoning if the page has changed.

Watch-and-learn

BackBrain's demonstration mode — you perform the task once while it observes without acting; it summarizes what it understood and saves only when you say "save".

MCP (Model Context Protocol)

The open industry standard for connecting AI agents to external tools; BackBrain can add tools from the official MCP registry in one click.

Connector recipe

A small JSON description that turns any REST API into tools the agent can use — no programming.

Tool pinning

A defense that refuses a connector whose tools changed after you approved it, until you review and re-approve.

Purchase mandate

A pre-authorization (merchant, maximum amount, expiry) that must exist before the agent can buy anything — on top of the normal approval prompt.

Sentinel trigger

A rule you define — such as "a new file matching this pattern appears" — that can start a task automatically.

Routine

A task scheduled to run daily or weekly in the background.

Kill-switch

The one-click control that immediately stops all background and autonomous activity.

Prompt injection

An attack where text on a webpage or in a tool's output tries to give the agent instructions; BackBrain screens and scrubs external content before it can influence the agent.

Helper agents

Parallel sub-agents (research, planning, verification, extraction) that assist the main run; they can read and think but never touch the browser session or credentials.