Glossary

Account sign-in

Logging an agent into an account you already have with a company, such as a ChatGPT or Claude plan, instead of pasting an API key. Some companies allow this only in their own apps, so using it elsewhere can break their terms.

Also called: OAuth

Agent

A program that uses a model to plan and carry out tasks on its own: running commands, reading and writing files, and browsing the web. A chatbot answers you; an agent acts for you.

Allow list

A list of the only people, chat accounts, or websites an agent will accept. Anything not on the list is refused.

Also called: allowlist, allow-list

API

The way one program talks directly to another company's service, with no website in between. Agents reach cloud models through an API.

API key

A long secret code from a provider that lets an agent use its API and bills the usage to your account. Treat it like a password, because anyone who has it can spend your money.

Also called: key

Approval

A pause where the agent shows what it is about to do and waits for your yes or no. Approval controls what runs, but anything you approve runs with your own permissions.

Also called: approvals

Ban risk

The chance that a company closes or suspends your account because you used it in a way its terms forbid.

Chat bot

An account in a chat app, such as Telegram, that passes your messages to the agent and its replies back to you. Unless you limit it with an allow list, anyone who can message the bot can give your agent orders.

Also called: bot

Chat plan

A flat monthly plan for a company's own chat app, such as ChatGPT Plus or Claude Pro. It is a separate product from pay-as-you-go API access.

Also called: subscription

Cloud model

A model that runs on a provider's computers instead of yours. Your prompts, and any files the agent sends along, leave your computer.

Command

An instruction typed into a terminal that tells the computer to do something, such as install a program or delete a file. Agents run commands too, which is where most of their power and most of their risk come from.

Container

A sealed-off space on your computer, with its own files and programs, where an agent can work without touching your real files. Docker is the usual way to make one.

Docker

Free software that creates and runs containers. Docker Desktop is the version with a normal app window for Windows and Mac.

Free tier

The amount of use a provider gives away at no charge. It usually comes with lower limits, and some providers use free-tier prompts to improve their products.

Frontier model

One of the newest and most capable models from the leading AI companies, such as Claude Opus 5.5.

Gateway

The part of an agent that keeps running in the background so it can receive chat messages and run scheduled jobs. The computer it runs on has to stay on.

Host

Your real computer, as opposed to a container running inside it. Host access means letting the agent out of its container onto your real files and programs.

Also called: host access

Local model

A model that runs on your own computer instead of a provider's. It costs nothing per use and keeps your prompts on your machine, but it needs a lot of RAM and is weaker than the best cloud models.

Localhost

The address a computer uses to mean itself. A page at http://localhost comes from your own computer and is not open to the internet.

MCP

Model Context Protocol: a standard way to plug outside tools into an agent, where each MCP server is a program that offers one or more tools. An MCP server gives your agent its tools, and a server that runs on your computer can reach whatever your account can, so add only servers you trust.

Also called: MCP server

Model

The AI system that reads your request and decides what to do next, such as Claude Opus 5.5 or Gemini. The agent is the program, the model is the reasoning inside it, and most agents let you swap one model for another.

Ollama

A free app that downloads and runs local models. LM Studio does the same job with more of a point-and-click window.

Also called: LM Studio

Open source

Software whose code is published so anyone can read, use, and change it, usually at no cost.

Pay-as-you-go

Billing where you pay for exactly what you use, measured in tokens, instead of a flat monthly price.

Persona

The name, tone, and speaking style you ask an agent to use. It shapes how the agent talks to you, not which tools it is allowed to use.

Port

A numbered door on a computer that a program listens on. In http://localhost:5080, the port is 5080.

Prompt

The text a model receives: your request, plus any files or web pages the agent passes along with it.

Prompt injection

Instructions hidden in a web page, email, or file that try to make the agent do something you did not ask for. It is the main reason not to connect an agent to your email or bank.

Provider

A company that runs models and sells access to them, such as Google, Anthropic, OpenAI, Groq, or DeepSeek.

RAM

A computer's short-term working memory, measured in gigabytes (GB). A local model has to fit in it, so RAM decides how strong a local model you can run.

Rate limit

A cap on how many requests you can send in a minute or a day. When the agent hits it, the agent stalls until the limit resets.

Relay

A third-party service that sits between your agent and a provider and forwards your requests, often at a discount. It can read everything you send through it.

Also called: proxy

Revoke

Cancel a key, token, or account connection at the service that issued it, so an agent that holds it can no longer use it.

Sandbox

A wall, enforced by the operating system or a container, that limits what a program can reach. An approval prompt is not a sandbox: it asks, but it does not block.

Schedule

A task the agent runs by itself at set times, such as every morning. Scheduled jobs usually run with nobody watching, so nobody is there to answer an approval.

Also called: scheduled job, scheduled task

Skill

A saved set of instructions, or a small program, that an agent can reuse for one kind of task. Some agents can write new skills for themselves.

Spend limit

A setting in your provider account that stops charges once you reach a monthly amount. It is your backstop if an agent gets stuck repeating itself.

Also called: spend cap, spending limit

Sub-agent

A second agent that the main agent starts to handle one part of a bigger job. It works in its own conversation and reports its result back to the agent that started it.

Also called: subordinate agent, subagent

System prompt

The standing instructions an agent sends to the model before your message in every conversation. It sets the agent's role and how it uses its tools.

Terminal

A text window where you type commands. On Windows it is usually PowerShell; on a Mac it is the Terminal app.

Also called: PowerShell

Token

The unit providers use to measure and bill text. A token is roughly three-quarters of an English word.

Tool

Something an agent can use besides talking, such as a web browser, a file editor, or a terminal. Every tool is one more thing the agent can do for you, and one more thing that can go wrong.

Unofficial route

A way to use a model that its company does not support, such as signing a third-party agent into a chat plan or going through a relay. These can be cheap, but they carry ban, billing, or privacy risk.

Virtual machine

A whole simulated computer, with its own operating system, running inside your real one. It walls off its contents much like a container does, but it is heavier and more complete.

Also called: VM

Workspace

The folder where an agent keeps the files it reads and makes. On some agents it is only a starting point, not a wall.

WSL2

Windows Subsystem for Linux: a part of Windows that runs Linux inside it. Some agents, and some Docker setups, need it.

Also called: WSL