Glossary

A working vocabulary for this guide. When a term has a vendor-specific meaning that differs from common usage, both are noted.

Agent. A system that uses a model in a loop to plan, take actions through tools, observe results, and decide what to do next, until a goal is reached or a stop condition fires. In practice this guide treats anything that runs a model with tool use across more than one step as an "agent."

Tool / Tool use / Function calling. A structured way to let a model call code, an API, or a connector. ChatGPT and OpenAI APIs use the term "tool calling" / "function calling" (see the OpenAI docs). Claude uses "tool use" (see the Anthropic docs). Gemini uses "function calling" / "tools" (see the Gemini API docs).

MCP (Model Context Protocol). An open protocol — primarily driven by Anthropic and adopted by other vendors — for connecting LLM clients to data sources and tools through standardized servers. See the MCP specification.

MCP server. A process that exposes resources, prompts, and tools over MCP. Can be local (stdio) or remote (HTTP/SSE).

MCP client. A host application that talks to MCP servers. Examples include Claude Desktop, Claude Code, and various IDE plugins.

Connector (Claude). Anthropic's user-facing name for an integration. Custom connectors use remote MCP under the hood. See Claude support: connectors.

Project (Claude / ChatGPT). A workspace concept that bundles instructions, files, and conversation history. Distinct from a "Custom GPT" in ChatGPT. See Projects in ChatGPT and Anthropic's project surfaces.

Custom GPT (ChatGPT). A configurable variant of ChatGPT with custom instructions, knowledge files, and actions. Distinct from a "Project." See Creating and editing GPTs.

Custom instructions. Persistent instructions applied to all conversations on a given surface. ChatGPT, Claude, and Gemini all expose a version.

Codex CLI. OpenAI's local terminal coding agent. See https://developers.openai.com/codex/cli and https://github.com/openai/codex.

Codex (cloud). OpenAI's cloud-side coding agent surface, accessed from ChatGPT.

Agents SDK (OpenAI). A Python framework for building agents that orchestrate models, tools, handoffs, and guardrails. See https://github.com/openai/openai-agents-python.

Antigravity (Google). An agent-first IDE from Google built around Gemini. See Codelab.

Gemini CLI. Google's open-source command-line agent for Gemini. See https://github.com/google-gemini/gemini-cli.

Google AI Studio. Google's web playground and API key management surface for the Gemini API. See https://ai.google.dev/aistudio.

Computer use. A capability where a model controls a virtual computer (mouse, keyboard, screen). Anthropic's tool: computer-use-tool. OpenAI's tool: Computer-Using Agent.

browser-use (library). An open-source Python library for letting LLMs drive a real browser. See https://github.com/browser-use/browser-use.

GitHub Copilot cloud agent. A cloud agent that opens PRs against your repo, distinct from the in-editor Copilot. It was formerly called Copilot coding agent. See About GitHub Copilot cloud agent.

Eval / Eval set. A reproducible set of inputs + expected behaviors used to measure agent quality across changes.

Red team. A deliberate effort to find failure modes — prompt injection, data exfiltration, unsafe tool use — before users do.

Human-in-the-loop (HITL). A pattern where the agent pauses for human review before high-impact tool calls.

Drift risk. A subjective rating in this guide for how quickly a page is likely to go stale. Low = stable for ≥1 year. Medium = re-check quarterly. High = re-check monthly.