Reference

Glossary

Names you will keep seeing. Short first, then the useful detail.

Agent: A loop that plans and uses tools, not just a single reply.; A chat replies once. An agent can search, call tools, inspect results, and continue until a goal is met or a stop rule fires. Treat agents as drafts of action. Keep consequential clicks human.
API: A machine door. Programs talk here instead of a screen.; An application programming interface lets software send prompts and receive answers without a graphical chat box. Useful for builders. Easy to over-share if keys and logs are sloppy.
Connector: A signed-in bridge from a chat product to another service.; Gmail, Drive, GitHub, Notion, calendars. A connector is a product feature, often built on MCP or a similar protocol. It has scopes: read, write, send, delete. Start with the smallest useful grant.
Context window: How much text the model can see at once.; Measured in tokens. Includes the system instructions, your messages, attached files, tool results, and the model's own replies. When it fills, older turns are dropped or summarized. A long chat is not infinite memory.
Custom GPT / Gem / Project: A saved setup: instructions, files, sometimes tools.; Names differ by product. The idea is the same: reuse a brief. This is not live access to your inbox unless you also connect a tool. Instructions without data still hallucinate.
GUI: Graphical user interface: buttons, menus, text boxes.; The friendly screen. ChatGPT, Claude, Gemini, Grok, Copilot in a browser or app. The GUI does not tell you which model, data controls, or connectors are active. Check the product and plan.
Hallucination: A fluent answer that is not grounded in a source.; Models complete likely text. They do not automatically know when they are inventing. Ask for sources, require “not stated” on missing facts, and check anything that will be sent, paid, or published.
Host (MCP): The AI application you type into.; Claude, ChatGPT, Grok, Cursor, VS Code. The host owns the conversation and contains one or more MCP clients that talk to servers.
MCP: Model Context Protocol. USB-C for AI applications.; An open standard, introduced by Anthropic in 2024 and later moved under the Linux Foundation, that lets any compatible assistant talk to tools and data through a shared protocol. Host, client, server. Build a server once; many clients can use it.
MCP server: A program that exposes tools, resources, or prompts.; The server never is the model. It offers named capabilities with schemas. The model proposes a call; the host and your permissions decide whether it runs.
Model: The engine that predicts the next tokens.; GPT, Claude, Gemini, Grok, Llama, Mistral, Qwen, DeepSeek. A product may switch models without changing its name. Always look at the model picker when the work is sensitive or expensive.
Product: The place you open: app, site, or desktop tool.; ChatGPT, Claude, Grok, Gemini, Copilot, Perplexity, Cursor, OpenCode. The product decides logging, plans, connectors, and which models you can reach.
Prompt: The instruction you give the AI.; Not magic language. A good prompt names the source, the task, a boundary, and the output shape. Then you iterate. One perfect paragraph is rarer than three short rounds.
RAG: Retrieval-augmented generation: look up, then answer.; The system fetches passages from files or a database and stuffs them into the prompt. Better than hoping the model memorized your handbook. Still depends on the retrieved chunks being the right ones.
Scope: The permission a connector is allowed to use.; Read-only mail is not send. Search Drive is not share. Treat scopes like keys. Review them the way you would review an app's photo access.
Skill: A reusable instruction pack the assistant can load.; Skills tell the model how to work: voice, format, process. They do not, by themselves, open Gmail. Pair a skill with a connector when the work needs live data.
System prompt: Hidden instructions the product adds before your text.; Safety rules, tool lists, personality, date. You do not always see them. They still consume context and steer behavior.
Token: A chunk of text the model reads and writes.; Roughly 3–4 characters of English, or about 0.75 words. Limits, bills, and context windows are counted in tokens, not pages.
Tool: A capability the model can invoke during a reply.; Web search, code execution, image generation, a calendar lookup. Built-in tools live inside the product. MCP tools arrive through a server. Either way, the model proposes; policy and you decide.
Weights: The downloaded numbers that are the model.; Open-weight means you can download the file. The license may still limit use. Running weights locally keeps the prompt on the machine only if remote fallbacks, telemetry, and extensions are actually off.