Glossary
One hundred short, plain definitions of the terms you meet when coding agents do real work: the Model Context Protocol and its pieces, how agents sign in safely, the rules files they read, the agents themselves, how several people share one page, and how AI search finds and cites what you publish. Each term has a sketch of how it works, numbered steps and its source.
Start with the group that matches your question. Setting up an agent? MCP explains the protocol your agent speaks and Sign-in and safety explains what happens when it signs in. Writing a CLAUDE.md or AGENTS.md? Rules and context. Deciding what an agent should do on its own? Coding agents. Several people's agents editing one document? Working together. Want your pages read and cited by ChatGPT, Claude and Google's AI Overviews? AI search.
Every entry has the same shape: the answer in two or three sentences, a sketch of how it works with numbered steps, then the detail that matters in practice, with its source. When a term is one an agent actually uses on a board, the entry says how, and links the matching setup guide or comparison.
MCP
- What is JSON-RPC?JSON-RPC 2.0 is the small remote-call format under every MCP message: a request with an id, a reply echoing it, a notification with none, and what bites.
- What is a local MCP server?A local MCP server is a program your AI client starts on your own computer and talks to over stdin and stdout. What it can reach, and when to pick one.
- What is MCP?MCP, the Model Context Protocol, is the open standard that connects AI apps to tools and data. What it stands for, how it works, and what it is not.
- What are MCP Apps?MCP Apps is the MCP extension that lets a server return an HTML interface the chat renders in a sandboxed iframe. How it links to tools and talks back.
- What is an MCP client?An MCP client is the connector inside an AI app that talks to one MCP server. How it differs from the host, what it sends, and where you run into it.
- What is an MCP connector?An MCP connector is a remote MCP server attached to your Claude account. How to add one, how it signs in, where it connects from, and what to check first.
- What is MCP elicitation?MCP elicitation lets a server ask the person a question mid-request: a form the client renders, or a URL for secrets. Modes, schema limits, retries.
- What is an MCP gateway?An MCP gateway puts one URL and one sign-in in front of many MCP servers. How it routes tool calls, holds credentials, and where the spec draws the line.
- What is an MCP host?An MCP host is the app that holds MCP clients, the model and your approvals: Claude Code, Cursor, a chat UI. What it does and why trust lives there.
- What is the MCP Inspector?The MCP Inspector is the official tool for poking at an MCP server: list its tools, call one, read resources and watch the protocol, in a browser or CLI.
- What is MCP OAuth?MCP OAuth is how an AI agent signs in to a remote MCP server: OAuth 2.1 with PKCE, found through metadata, with the person approving in a browser.
- What are MCP prompts?MCP prompts are message templates a server offers and a person selects, often as slash commands. What prompts/get returns and when to use one.
- What are MCP resources?MCP resources are URI-addressed data an MCP server exposes as model context. How listing, reading, templates and updates work, and when a tool fits better.
- What is MCP sampling?MCP sampling lets a server ask the client's model for a completion without holding a key. How it flows, what the person approves, and why it is deprecated.
- What is MCP security?MCP security covers who an agent connects to, what its tools may do and which tokens go where: the attacks the spec names and the controls that stop them.
- What is an MCP server?An MCP server is a program that exposes tools, resources and prompts to AI apps over the Model Context Protocol. How it works, an example, and pitfalls.
- What are MCP tools?MCP tools are functions a server exposes for a model to call: name, description, JSON Schema in, content out. How they differ from resources and prompts.
- MCP vs A2A: what's the difference?MCP connects an agent to its tools and data. A2A connects agents to each other. What each protocol defines, how they fit together, and when you need both.
- MCP vs API: what's the difference?An API is a contract your code calls; MCP is a protocol a model discovers at run time. How they differ, where each wins, and why MCP servers wrap an API.
- MCP vs RAG: what's the difference?RAG fetches text into the prompt before the model answers. MCP lets the model call tools and read sources while it works. How they differ and combine.
- What is a remote MCP server?A remote MCP server runs on the internet, not on your laptop. Your agent reaches it at a URL over Streamable HTTP and usually signs in with OAuth.
- What is Streamable HTTP in MCP?Streamable HTTP is how MCP runs over the network: one endpoint, a POST per message, JSON or SSE back. Versus stdio and HTTP+SSE, and what changed in 2026.
Sign-in and safety
- What is a bearer token?A bearer token is a credential that works for whoever holds it. Where to put it, what 401 and 403 mean, and how agents keep one safe.
- What is the confused deputy problem?The confused deputy problem: a trusted program spends its own authority for someone who should not have it. How it shows up in MCP proxies and AI agents.
- What is the device code flow?The device code flow (RFC 8628) signs in a machine with no browser: it shows a short code, you approve on your phone or laptop, and it polls for the token.
- What is dynamic client registration?Dynamic client registration (RFC 7591) lets an app get an OAuth client ID by posting its details to a server it has never met, with no form to fill in.
- What is human in the loop?Human in the loop means a person approves what an automated system does before it takes effect. Where the checkpoint goes for AI agents, and what it costs.
- What is OAuth 2.1?OAuth 2.1 folds a decade of OAuth 2.0 fixes into one spec: PKCE on every code flow, exact redirect matching, no implicit grant. What it means for agents.
- What is PKCE?PKCE (RFC 7636) stops a stolen authorization code from being cashed in. How the verifier and challenge work, why S256, and why MCP clients must use it.
- What is prompt injection?Prompt injection is when text an AI model reads changes what it does. The direct and indirect kinds, why filters fall short, and what limits the damage.
Rules and context
- What is agent memory?Agent memory is what an AI agent keeps after its context window is gone: instruction files, its own notes, memory tools. How Claude Code memory works.
- What is AGENTS.md?AGENTS.md is a Markdown file at the root of a repo that tells coding agents how to build, test and work in it. Where it goes, what goes in, who reads it.
- What is CLAUDE.md?CLAUDE.md is the file Claude Code reads at the start of every session: your project's commands, conventions and rules. Where it lives, load order, imports.
- What is context engineering?Context engineering is deciding what goes into a model's window at each step: instructions, tools, history, data. How it differs from prompt engineering.
- What is context rot?Context rot is the drop in accuracy as a model's input grows, even on simple tasks. What Chroma's 18-model study found, and how to keep an agent sharp.
- What is a context window?A context window is everything a model can read while answering: instructions, tools, history and its reply. What fills it and what happens at the limit.
- What are Copilot instructions?Copilot instructions are Markdown files in .github that Copilot adds to every request: the repository-wide file, path-specific files, priority order.
- What are Cursor rules?Cursor rules are .mdc files in .cursor/rules that put standing instructions into the Agent's context: where they live, the four types, what gets skipped.
- What is GEMINI.md?GEMINI.md is Gemini CLI's instruction file, read from your home folder, the project and its parents, and sent with every prompt: where it lives, imports.
- What is prompt caching?Prompt caching lets a model API reuse the unchanged start of a prompt across requests, cutting cost and latency. How prefixes match, expire and break.
- What is a system prompt?A system prompt is the text an application gives a model before any user message: role, rules, context. How Claude and OpenAI take it and what it can't do.
- What are Windsurf rules?Windsurf rules are Markdown files Cascade reads on every message or on demand: global_rules.md, .windsurf/rules, trigger modes, size limits and memories.
Coding agents
- What is the A2A protocol?The A2A protocol lets one AI agent find another, hand it a task and get results back over HTTP. Agent Cards, task states, transports, and A2A versus MCP.
- What is ACP?ACP is Zed's Agent Client Protocol: JSON-RPC over stdio that lets editors like Zed and JetBrains run coding agents. IBM's unrelated ACP is now part of A2A.
- What is an agent loop?An agent loop is the cycle an AI agent runs: the model picks a tool, your code runs it, the result goes back, and it repeats until the model stops asking.
- What is an agent sandbox?An agent sandbox is the boundary an AI agent's commands run inside, limiting files and network. OS-level sandboxes in Claude Code and Codex, and cloud VMs.
- What is an agent SDK?An agent SDK is a library that runs the agent loop for you: tools, permissions, context, sessions. What Claude's and OpenAI's give you, and when to use it.
- What are agent skills?Agent skills are folders with a SKILL.md that an AI agent loads only when a task calls for it. The format, how loading works, and what to put in one.
- What is agentic AI?Agentic AI is the umbrella term for systems where a model decides the steps. What it means, how vendors stretch it, and how to tell a real one from hype.
- Agentic AI vs AI agent: what's the difference?Agentic AI is the property and the field; an AI agent is one program that has it. The distinction side by side, with a diagram and which word to use when.
- What is agentic coding?Agentic coding is the practice of handing a coding agent a task, letting it read, edit and run, then reviewing the result. The habits and the tools.
- What is agentic engineering?Agentic engineering is Karpathy's name for professional work where you direct coding agents and review their output instead of typing the code yourself.
- What is an agentic workflow?An agentic workflow lets an AI model plan, use tools, check its own work and revise instead of answering in one pass. Ng's four patterns, with an example.
- What is an AI agent?An AI agent is a model that uses tools in a loop to finish a task for you. What counts as one, how the loop runs, what it is not, and real examples.
- What is AI code review?AI code review is an agent reading a pull request and commenting where it finds bugs. What Codex, Copilot, Bugbot and Claude Code do, and what stays yours.
- What is AI pair programming?AI pair programming: a model in your editor suggests the next line or function and you keep control. Where it came from and how it differs from agents.
- What is a browser agent?A browser agent reads web pages, clicks, types and fills forms in your own Chrome or a hosted browser. What it controls and the guardrails that matter.
- What are Claude Code hooks?Claude Code hooks run your commands, HTTP calls or prompts at points in a session, such as before a tool runs, and can add context or block the action.
- What is Claude Code plan mode?Claude Code plan mode: Claude reads your code and proposes a plan while edits stay blocked until you approve. How to turn it on, what still runs, and why.
- What are Claude Code plugins?A Claude Code plugin is a directory of skills, subagents, hooks and MCP servers installed as one unit from a marketplace: what is inside, when to make one.
- What are Claude Code slash commands?Claude Code slash commands start a message with / to run a built-in action or a saved prompt. Where they live, how arguments work, and what goes wrong.
- What is a cloud agent?A cloud agent runs a coding task in a hosted VM, not on your laptop: it clones the repo, works while you are away and pushes a branch for you to review.
- What is a coding agent?A coding agent is an AI agent whose tools are your repository: it reads files, edits them, runs commands and tests, and hands you a diff to review.
- What is computer use?Computer use lets an AI model drive a desktop through screenshots, clicks and keystrokes. How Anthropic's tool loop works, what runs where, and its limits.
- What is function calling?Function calling lets a model ask your program to run a function with JSON arguments and read the result. How it works in the OpenAI and Claude APIs.
- What is a multi-agent system?A multi-agent system splits one task across several AI agents: a lead that delegates to parallel workers. What it buys, what it costs, when not to use it.
- What is a ReAct agent?A ReAct agent interleaves thought, action and observation: reason about the next step, call a tool, read the result, think again. Yao et al., 2022.
- What is spec-driven development?Spec-driven development has you write what to build and why first; a coding agent then turns that spec into a plan, tasks and code checked against it.
- What is a subagent?A subagent is a separate agent your coding agent starts for one task. It gets its own context window, prompt and tools, and sends back only a summary.
- What is vibe coding?Vibe coding means asking an AI to write code you never read. What Karpathy meant by it, where it works, and why unreviewed code is the catch.
- What is YOLO mode?YOLO mode turns off every approval prompt so a coding agent runs unasked. What Gemini CLI, Cursor and Claude Code call it, and when it's safe to use.
Working together
- What is an AI artifact?An AI artifact is a standalone piece of work an assistant builds beside the chat: a document, page, app or diagram you edit in place and share afterwards.
- What is an audit trail?An audit trail is a chronological record of who did what and when, kept so events can be reconstructed. What belongs in one, and what makes it useless.
- What is ChatGPT Canvas?ChatGPT Canvas is a side-by-side workspace where you and ChatGPT edit one draft or code file in place, with shortcuts for length, polish, bugs and ports.
- What is content-addressable storage?Content-addressable storage names each file by a hash of its bytes, so identical content is stored once and any copy can be verified. With Git and IPFS.
- What is a CRDT?A CRDT is a data type whose copies can each change on their own and still end up equal when they sync, with no server and no conflicts resolved by hand.
- What are design tokens?Design tokens are named design values (colors, spacing, type) kept in one file so tools and code agree. The W3C format, aliases, and how agents use them.
- What is Gemini Canvas?Gemini Canvas is the Gemini app's side panel for docs, slides, apps and code: you edit beside the chat, step through saved versions, then export or share.
- What is a living document?A living document is kept current for as long as it is used instead of being finished and filed. What it needs: one link, open editing and a full history.
- What is a merge conflict?A merge conflict is two edits to the same lines that a tool cannot combine. How git marks it, how to resolve it, and how to make it rare.
- What is operational transformation?Operational transformation (OT) shifts each incoming edit by the edits it crossed so every editor's copy converges. How Google Docs does it, and its limit.
- What is optimistic concurrency control?Optimistic concurrency control: write only if the version you read is still current, otherwise retry. Where it beats locks and where it fails.
- What is a rebase?A rebase replays a change made against an older version on top of the newest one: what git rebase does, where it stops, and how an agent rebases a publish.
- What is a shared workspace?A shared workspace is one place where a team sees and edits the same things. What it needs, how the new kind adds coding agents, and where teams go wrong.
- What is a single source of truth?What a single source of truth is, why copies drift, and how to keep one page the place a fact is edited while everything else reads from it.
- What is a three-way merge?A three-way merge compares two edited copies with the base they came from, so it knows which side changed each line. How git and diff3 do it.
- What is version history?Version history keeps every saved state of a file with who changed it and when. What it records, how restoring works, and where tools differ.
AI search
- What are AI citations?AI citations are the links ChatGPT, Perplexity and AI Overviews attach to an answer. How each engine shows them, how pages get picked, how to earn one.
- What are AI crawlers?AI crawlers are the bots OpenAI, Anthropic, Perplexity and Google send to fetch pages for training, search and live answers. Names, jobs, robots.txt rules.
- What are AI Overviews?AI Overviews are Google's generated summaries above search results, with source links. What triggers them, how pages get in, and the controls you have.
- What is AI search optimization?AI search optimization is what a site does so AI search engines can crawl, read and quote it: robots.txt, sitemaps, llms.txt, JSON-LD, plain answers.
- What is answer engine optimization?Answer engine optimization (AEO) is writing a page so an AI Overview, snippet or assistant can lift your answer whole. What changes on the page, and why.
- What is generative engine optimization?Generative engine optimization (GEO) is making your pages the ones an AI-written answer cites. What the 2023 paper measured, what worked, what did nothing.
- What is llms.txt?llms.txt is a proposed Markdown file at /llms.txt that gives AI agents a curated map of a site. What goes in it, who publishes one, and what is unproven.
- What does machine-readable mean?Machine-readable data can be processed by a computer without a person. For a website that means real HTML, JSON-LD, a sitemap and llms.txt for AI agents.
- What is programmatic SEO?Programmatic SEO builds many pages from data and a template. The version that works, the thin version Google penalises, the checks that tell them apart.
- What is semantic HTML?Semantic HTML uses main, article, nav and headings for what content is, not how it looks, so screen readers, crawlers and AI agents can read and edit it.
- What is structured data?Structured data is typed, machine-readable markup on a page. JSON-LD, FAQPage, BreadcrumbList and DefinedTerm explained, and what changed for FAQ schema.
- What is topical authority?Topical authority is depth on one subject: a hub plus a cluster of specific, linked pages. How to build one, and how it differs from domain authority.
- What is zero-click search?A zero-click search ends on the results page. What that means for a site that wants to be the quoted answer, with Pew's 2025 numbers and Google's reply.
FAQ
- How is this glossary organised?
- Six groups: MCP, Sign-in and safety, Rules and context, Coding agents, Working together, and AI search. Each entry opens with a direct answer, then a sketch of how the thing works with numbered steps, then examples, common mistakes and related terms.
- Where do the glossary definitions come from?
- Each entry names its primary source: the MCP specification, an RFC, a vendor's documentation or a paper, read on the date shown on the page. Where a term touches Intraface, the entry uses the product as a worked example and says so.
Stop sending HTMLover Slack.
Create an account in a minute, connect your coding agent, and make your first board.