Concepts
Glossary
Plain-language definitions for the computer, AI, and AFK terms used in these docs.
New to command-line tools? Start with the first section, then look up unfamiliar words as you need them. Each term links to a guide with more detail.
Using your computer
| Term | Meaning in these docs |
|---|---|
| Terminal | An app where you type instructions to your computer and read its replies as text. You run the afk command here. |
| Shell | The program inside a terminal that reads and runs your commands, such as zsh, Bash, or PowerShell. Commands written for one shell may need different wording in another. |
| PowerShell | A shell commonly used on Windows. AFK can run on native Windows and uses PowerShell for agent shell commands when Git Bash is unavailable. |
| Command | An instruction you type and submit with Enter, such as afk --version. Commands beginning with afk go in your terminal; AFK's slash commands, such as /help, go inside its conversation. |
| CLI | Short for command-line interface: a way to use software by typing commands rather than clicking buttons. The installed AFK CLI is named afk. |
| Node.js | The software that runs AFK on your computer. Install a supported version before installing AFK. |
| npm | A package manager: a tool that downloads and installs software made for Node.js. You use it to install the package named agent-afk, which provides the command named afk. |
| PATH | A list of folders your shell searches when you type a command. AFK can be installed but still not found if its command's folder is missing from this list. |
| Working directory | The folder a program is currently working from. For AFK, choose the folder containing the files you want it to work on. |
| Git repository | A project folder whose file history is tracked by Git. AFK's worktree features need a Git repository, but an ordinary conversation does not. |
| Worktree | A separate working copy of a Git project, so an agent can edit files without mixing them with another task's edits. It separates files, not all access to your computer. |
| Environment variable | A named setting a program reads when it starts, often used for configuration or credentials. AFK also loads saved environment settings from its user configuration file. |
Models, sign-in, and cost
| Term | Meaning in these docs |
|---|---|
| Provider | The service or connection AFK uses to reach an AI model, such as Anthropic, OpenAI, xAI, or a compatible local server. The provider determines which credentials and connection settings are needed. |
| Model | The AI system that reads your request and generates a response or decides which tools to use. You can choose different models for different tasks. |
| Model slot | A configurable nickname such as small, medium, or large that points to a model and can include its connection settings. It is not a separate subscription. |
| API key | A secret issued by a provider that lets AFK use that provider's service on your account. Keep it private and check the provider's billing arrangements before using it. |
| OAuth token | A secret sign-in credential that gives software access to an account without sharing the account password. AFK supports several provider-specific token sources, and tokens can expire. |
| Token | In model usage, a small piece of text counted toward input, output, and often cost; it is not necessarily a whole word. A sign-in token or Telegram bot token means a secret credential instead. |
| Local model | An AI model served by software you run yourself, rather than a hosted provider account. AFK connects to that server; installing AFK alone does not install or start a local model. |
Conversations and unattended work
| Term | Meaning in these docs |
|---|---|
| REPL | Short for read-evaluate-print loop: AFK's ongoing conversation in your terminal. Type a request, read the response, then continue in the same session. |
| Session | One conversation together with its model choice, permissions, and recorded activity. A session can contain many messages. |
| Turn | One exchange starting with a new message, during which the agent may respond and use tools. It is part of a session, not necessarily a new conversation. |
| Context | The information available to the model for its current response, including instructions, conversation, and tool results. The model's context window limits how much can fit at once. |
| Compaction | Summarizing older conversation to make room in the model's context while keeping recent turns. It is not the same as deleting all saved conversation records. |
| Tool | An action the agent can call, such as reading a file or running a command. Available tools and permissions determine what it can actually do, beyond writing a reply. |
| Subagent | A separate agent conversation given a smaller task by the main agent. It returns findings to the agent that sent it and can have different tool access or a different model. |
| Daemon | AFK's program for running scheduled or queued work without an interactive conversation open. The daemon still needs to be running on a computer to do that work. |
| Service | A program supervised by your operating system so it can start automatically and restart after a crash. AFK can install its daemon or Telegram bot this way. |
| Permission mode | A setting that controls checks on agent actions, such as approval for file access outside the working folder or read-only planning. It is not a guarantee that every shell command asks permission or that the agent is sandboxed. |
Extensions
| Term | Meaning in these docs |
|---|---|
| Skill | A reusable procedure or workflow you can ask AFK to run. It can give instructions to the current conversation or run in a separate agent session. |
| Plugin | An installable package of extensions, such as skills, agent definitions, hooks, or tool connections. Some plugins run code, so install only ones you trust. |
| MCP | Short for Model Context Protocol: a shared way for external programs or services to offer tools to AFK. An MCP server is the program or service providing those tools. |
| Hook | Code that runs at a defined point, such as before the agent uses a tool. Hooks can check, change, or block actions where the runtime supports it. |
| AFK.md | A file of instructions and project information that AFK can load into its conversations. Instructions guide the model; they are not enforced access controls. |