Glossary
AI website building, defined
Short, plain-English definitions of the terms you will meet when building websites with AI — no jargon required to get started.
BYOK (bring your own key)
A model where a tool runs on an API key you obtain directly from an AI provider, instead of reselling AI usage. You pay the provider's raw per-token rates; the tool adds no markup and needs no subscription. OpenThorn is a BYOK website builder.
AI website builder
A tool that turns a plain-language description into a working website. Modern builders use an AI agent that plans the site, writes real code, previews it, and deploys it — rather than filling templates.
AI agent
An AI system that works toward a goal in steps: it plans, calls tools (like writing files or running a compiler), checks the results, and keeps iterating until the task is done. OpenThorn's agent builds websites this way.
LLM (large language model)
The AI model that powers text and code generation — for example GPT (OpenAI), Claude (Anthropic), or Gemini (Google). Accessed programmatically through an API, billed per token.
API key
A secret credential from an AI provider that lets software call its models on your account. You create one in the provider's console, set spend limits there, and can revoke it at any time.
Token
The unit AI providers bill by — roughly three-quarters of a word of text or code. Providers price input and output tokens separately, usually per million tokens. Generating a typical website costs cents to a few dollars in tokens.
Context window
The maximum amount of text (in tokens) a model can consider at once — the conversation, instructions, and code it can 'see'. Larger windows let an agent work on bigger projects without losing track.
System prompt
Standing instructions given to a model before any user input — defining its role, rules, and tools. An AI website builder's system prompt is a large part of what makes its agent reliable.
Prompt
The instruction you give an AI model in plain language — a question, a description, or a task. In OpenThorn, describing the website you want is the prompt; everything the agent builds follows from it.
Inference
The act of running a trained AI model to produce output — text, code, or images. Providers bill inference by the token, which is why costs scale with how much context you send and how much the model generates.
Temperature
A model setting that controls randomness: lower values make output more predictable and consistent, higher values make it more varied and creative. For code generation, lower temperatures usually produce more reliable results.
Streaming
Delivering model output token by token as it is generated instead of all at once at the end. Streaming makes AI tools feel faster because you watch the answer — and the generated site — take shape live.
Rate limit
A provider-side cap on how many requests or tokens you can send per minute or day. Hitting a rate limit returns an error (usually HTTP 429); well-built tools retry automatically with backoff instead of failing your run.
Reasoning model
A model that spends extra compute 'thinking through' a problem step by step before answering — stronger on complex planning and debugging, but slower and more expensive per task than standard models.
Multimodal model
A model that understands more than text — typically images as well. Vision lets an AI website builder look at a screenshot of your site and spot visual problems, not just read the code behind them.
Tool calling (function calling)
A model capability where the AI asks the surrounding software to run a defined action — write a file, search the codebase, deploy a site — and uses the result to continue. Tool calling is what turns a chatbot into an agent.
Structured output
Model output constrained to a machine-readable format, usually JSON matching a schema. Builders rely on it so tool calls parse reliably instead of breaking on free-form prose.
Hallucination
When a model states something invented with confidence — a made-up library, a function that doesn't exist. Agents counter hallucinations by compiling, testing, and verifying output against reality before showing it to you.
Prompt caching
Reusing the unchanged prefix of a conversation between requests, so repeated context isn't billed or processed at full price. Long agent sessions get cheaper and faster when a provider supports caching.
Thinking budget
A configurable limit on how much internal reasoning a reasoning model may do before answering. Higher budgets help with hard problems but cost more tokens; builders tune it per provider and per task type.
Frontend vs full-stack
Frontend is what visitors see — layout, styling, interactions. Full-stack adds the server side: database, authentication, APIs. OpenThorn builds both — describe the app and the agent can wire up a database and sign-in flows, not just pages.
API endpoint
A specific URL a program calls to talk to a service — for example an AI provider's chat endpoint or your deployed app's backend. Each endpoint has defined inputs, outputs, and authentication.
Environment variable
A named configuration value (like a URL or a secret) provided by the environment instead of being written into code. Secrets such as signing keys belong in environment variables, never committed to the repository.
Dependency (npm package)
Reusable third-party code your project imports, installed via npm. Generated projects keep dependencies minimal and allowlisted so the bundle stays fast and the supply chain stays small.
Bundler
A build tool that combines source files into optimized assets browsers can load — resolving imports, tree-shaking dead code, minifying. OpenThorn's preview bundles generated code with esbuild compiled to WebAssembly, entirely inside your browser.
WebAssembly (WASM)
A low-level binary format that runs near-native-speed code in the browser. It's what lets heavyweight developer tools run client-side during preview instead of round-tripping every change through a server.
Single-page application (SPA)
An app that loads once in the browser and re-renders content as you navigate, without full page reloads. SPAs need prerendering so crawlers without JavaScript still see real content at each URL.
Hydration
The step where the browser takes pre-rendered HTML and attaches live JavaScript behavior to it, turning static markup into a working app without re-rendering from scratch. Done well: instant content, then interactivity.
Static prerendering
Generating the final HTML for each page at build time, so the server delivers complete documents instantly and crawlers can read everything without executing JavaScript.
Open Graph (OG) image
The preview picture shown when a link is shared on social media or in messengers, declared via og:image meta tags. Distinct OG images per page make shared links more clickable.
Structured data (schema.org)
JSON-LD annotations in the HTML that describe what a page is — product, FAQ, article, breadcrumbs. Search engines use it for rich results; answer engines use it to cite facts precisely.
Canonical URL
A <link rel="canonical"> tag telling search engines which URL is the definitive version of a page when several URLs could serve the same content, preventing duplicate-content dilution.
Sitemap.xml
A machine-readable list of all indexable pages with their last-modified dates, submitted to search engines so they discover new pages and re-crawl changed ones efficiently.
Robots.txt
A text file at the site root telling crawlers which paths they may fetch — increasingly used not just for Googlebot but to explicitly allow or restrict AI crawlers like GPTBot or ClaudeBot.
Core Web Vitals
Google's user-experience metrics — loading speed (LCP), interactivity (INP), and visual stability (CLS). Fast, stable pages rank better and convert more visitors.
Deployment target
Where a finished website gets published — hosting like Cloudflare Pages, Vercel, or any static host. Owning the exported code means you can deploy anywhere, never just inside the builder you started with.
Database schema
The structure of a database — tables, columns, and relationships. When an AI builder 'wires up a database', it creates and applies a schema your app reads and writes through a secure client.
Row-level security (RLS)
Database rules that restrict each signed-in user to their own rows of data at the database layer itself, so even a bug in application code cannot leak other users' data.
Go deeper: what is a BYOK AI website builder? · how to get an API key · FAQ