Vibe Coding Glossary: Every Term Explained in Plain English

If you've started poking around vibe coding, you've probably run into a wall of jargon within the first five minutes — tokens, context windows, agentic loops, MCP servers. None of it is actually complicated once someone explains it without the buzzwords. This page is that explanation: short, plain-English definitions of the terms you'll actually run into, in the order you're likely to meet them.

If you want the fuller picture first, see what vibe coding actually is. This page is the companion reference — bookmark it and come back whenever a new term shows up.

Vibe coding

Building software by describing what you want in plain language and letting an AI agent write and run the actual code, instead of typing every line yourself. You steer by reacting to results — "make the button blue," "now add a login" — rather than reading and writing code line by line. It's a way of working, not a specific tool.

AI coding agent

A piece of software that takes your plain-language instructions, writes real code, creates and edits files, and runs commands on your machine to produce something that actually works — not just a code snippet in a chat window. meshcode, Claude Code, and Codex CLI are all examples. The word "agent" specifically means it can act (edit files, run terminal commands, install things), not just talk.

Coding agent vs. chatbot

A chatbot (like the free ChatGPT or Claude web apps) answers questions and shows you code as text — copying, pasting, and running it is on you. A coding agent goes further: it opens your project, edits the actual files, runs the code, sees the errors, and fixes them itself in a loop. Same underlying AI model in many cases, very different amount of manual work left for you.

Prompt

The instruction you type to the AI — a sentence or paragraph describing what you want done. "Add a dark mode toggle to the settings page" is a prompt. Vague prompts get vague results; specific prompts (what page, what behavior, what it should look like) get results closer to what you actually meant on the first try.

Token

The unit AI models read and write text in — roughly a chunk of a word, not a whole word. "Coding" might be one token; "uncomplicated" might be split into two or three. Every prompt you send and every line of code the model writes costs tokens, and providers charge by the token, which is why token price is the real driver of how much AI coding costs you day to day.

Context window

The amount of text — your prompt, the code the agent has read, its own past responses — the AI model can "see" at once while working. Once a task exceeds it, the model starts losing track of earlier details, the way you'd lose track of a conversation from an hour ago. Bigger context windows let an agent hold more of a real codebase in mind at once, which matters more as projects grow past a single file.

Agentic loop

The read-act-check cycle an AI coding agent runs on its own: it reads relevant code, makes an edit, runs the result, reads any error it produced, and edits again — repeating until the task is done or it needs your input. This loop is what separates an "agent" from a one-shot chatbot reply: it's not writing code once and stopping, it's iterating against real feedback like a human developer would.

MCP (Model Context Protocol)

An open standard that lets an AI agent connect to outside tools and data — your calendar, a browser, a database, a file system — through a common interface, instead of every tool needing its own custom integration. In practice, MCP is what lets a coding agent do things beyond writing code, like reading a webpage, controlling a browser, or checking a project board, using the same connection method across different apps and providers.

Autonomous agent

An AI agent that keeps working through multiple steps on its own — writing code, testing it, fixing what breaks — without you approving every single action along the way. The opposite is a fully manual, approve-every-step mode. Most coding agents let you dial this up or down: tight supervision for sensitive changes, more autonomy for routine ones.

Multi-agent

Running more than one AI agent or model at the same time, rather than one model per window or session. This matters practically because different models are better at different things (one might be faster and cheaper for routine edits, another stronger for tricky debugging), and a multi-agent setup lets you use several in parallel instead of switching tools and copy-pasting context between them.

Pay-as-you-go / top-up

A pricing model where you add funds to a balance and usage is deducted from it at cost, instead of paying a fixed amount every month whether you used the tool or not. The appeal for vibe coding specifically is that usage is naturally uneven — some weeks are heavy, some are idle — so a flat $20/month subscription can mean paying for weeks you didn't touch the tool at all.

How meshcode fits these terms together

meshcode is a native desktop AI coding agent (Mac and Windows) built around several of these terms directly: it runs an agentic loop against your actual project files, supports MCP so it can reach beyond code into browsers and other tools, and runs multiple models at once instead of locking you into one model per window. It's free to start, and pricing is pay-as-you-go — top up from $1, no subscription, no monthly fee, no seat pricing, with a 5% + $0.50 flat fee taken only when you top up. If you already pay for Claude Code or Codex CLI, you can bring those keys into meshcode and use them at no extra meshcode token charge, since those providers bill you directly.

Now that the vocabulary is out of the way, the ideas behind it are simpler than the jargon made them sound — the best way to make them click is to try one prompt yourself and watch the agentic loop run.

👉 Download meshcode — Mac, Windows