Glossary

Plain-language definitions of the terms used throughout this documentation.

Agent : The system that carries out your request — it talks to a model, uses tools, and works through a task step by step.

Agent loop : The repeating cycle of ask the model → run a tool → feed back the result, which continues until the task is done.

Approval : Your decision to allow or deny an action the agent wants to take. See Permissions.

Auto-approve : A mode where a run approves its own actions, for unattended use in trusted environments.

Background job : A long-running command moved off the main conversation so work continues; its result arrives when it finishes.

Coordinator : A profile (or, in a swarm, a role) that manages several subagents working in parallel.

Endpoint : The URL where your model runs, which gdc connects to.

Fail closed : The safety principle that anything uncertain is denied, never allowed by default.

Headless : Running the agent without the interactive UI — via gdc print, for scripts and automation.

Hook : Your own code that runs in response to agent events, to audit, log, or block actions. See Hooks.

MCP server : An external program or service that adds tools and data the agent can use. See MCP servers.

Memory : Durable notes gdc keeps in view across sessions. See Memory.

Model / provider : The large language model gdc talks to, reached over an OpenAI-compatible endpoint. See Model providers.

Output style : A preset that changes how the agent phrases replies, without changing what it can do.

Permission : A rule that allows, asks about, or denies an action.

Plan mode : A way of running the agent so it researches and proposes a plan before changing anything. See Plan mode.

Plugin : A self-contained module that adds tools or hooks to gdc. See Plugins.

Profile : A preset that sets which tools are available and how the agent behaves. See Profiles & output styles.

Project : The working directory (and its subfolders) gdc scopes sessions, configuration, and memory to.

Server : A long-running gdc process (gdc serve) that remote clients connect to and that exposes a read-only API. See Admin guide.

Session : One saved conversation, including every message and tool run. See Sessions & projects.

Skill : A named instruction set the agent loads on demand. See Skills.

Subagent : A fresh, focused agent the main agent spins up for a self-contained piece of work. See Subagents & multi-agent.

Swarm : A set of gdc instances across machines that collaborate on tasks (experimental). See Swarm.

Tool : A concrete action the agent can take — read a file, run a command, search, and so on. See Tools.

Terminal app (TUI) : The interactive, full-screen program gdc-tui. See Interactive terminal.