Last updated: July 2026 · 32 terms across AI coding tools, agent architecture, and configuration
AI coding tools have developed their own vocabulary fast — much of it overlapping, some of it tool-specific, and a fair amount actively confusing (three different tools call three different things "rules," for instance). This glossary defines the terms that come up across the AI coding tool landscape, organized by category, with links to the deeper guides where relevant.
An AI system that can plan and execute multi-step tasks autonomously — reading files, writing code, running terminal commands, and correcting errors — rather than only responding to a single prompt with a single output. The term distinguishes tools like Cursor's Composer, Cline, and Claude Code from simpler autocomplete-only tools. See the AI Coding Tools Benchmark 2026 for how different agents perform on real multi-step tasks.
Real-time code suggestions that appear as you type, typically completing the current line or the next few lines. Distinct from agent capability — some tools (Cline, Aider) have no autocomplete at all and are agent-only, while others (GitHub Copilot in its original form) started as pure autocomplete tools before adding agent features.
A pricing and architecture model where the tool itself is free, and you provide your own API key from an AI provider (Anthropic, OpenAI, etc.), paying that provider directly at standard rates. Cline, Aider, and Continue.dev all use BYOK. The tool takes no markup on inference costs — see AI Coding Tools Pricing 2026 for real BYOK cost ranges.
The process by which an AI tool builds a searchable representation of an entire project — beyond just the currently open file — so it can retrieve relevant context automatically. Cursor's semantic indexing and Aider's repository map are two different approaches to this same underlying problem. Indexing quality is the main differentiator in how tools perform on large monorepos.
The maximum amount of text (measured in tokens) an AI model can process in a single request, including the prompt, any files provided, and the conversation history. A larger context window lets a tool consider more of your codebase at once. Claude Code's 1M-token context window (with Opus 4.7) is currently among the largest available in a coding agent.
Aider's term for a compressed, structural representation of a codebase — function signatures, class definitions, and file relationships — that fits within a model's context window without including full file contents. This lets Aider intelligently identify relevant files for a task on very large codebases. See the Aider Rules guide for how repo maps are generated and configured.
The basic unit AI models process text in — roughly ¾ of an English word on average. API costs for BYOK tools (Cline, Aider, Continue.dev) are calculated per token, which is why longer sessions and larger files cost more. See AI Coding Tools Pricing 2026 for real-world token cost examples.
A project-level file that gives an AI tool standing instructions about coding conventions — naming patterns, file structure, testing approach, error handling. Every major tool has its own format and filename: .cursorrules / .cursor/rules/*.mdc for Cursor, .clinerules for Cline, CONVENTIONS.md for Aider, CLAUDE.md for Claude Code, .windsurfrules for Windsurf, and .continue/rules/*.md for Continue.dev.
Instructions that define an AI agent's behavior, decision-making defaults, and universal constraints — distinct from a rules file, which defines coding conventions. A system prompt answers "how should the agent behave when uncertain?" while a rules file answers "how should the agent write code in this project?" See System Prompts for AI Coding Agents for the full distinction and templates for every major tool.
A frontmatter field used in Cursor's .mdc rules and Continue.dev's rules system that determines whether a rule loads on every single AI interaction (true) or only when relevant files are detected (false, typically paired with globs). Overusing alwaysApply: true across too many rules degrades instruction-following — see Cursor Rules Not Working for the specific failure pattern.
Wildcard file-path patterns (e.g. **/*.tsx, packages/api/**) used to scope a rule so it only activates for matching files. Used across Cursor's .mdc frontmatter, Continue.dev's config.yaml, and similar systems in other tools. Essential for monorepo configuration — see Best AI IDE for TypeScript Monorepo.
A .gitignore-syntax file that tells Aider which files and directories to exclude from its repository map and context — used to keep large monorepos or projects with generated files focused. Full reference in the Aider Rules guide.
Cline's term for global, cross-project behavioral configuration set in VS Code settings — functionally equivalent to a system prompt. Distinct from .clinerules, which is project-specific. See System Prompts for AI Coding Agents for how this maps across tools.
Cline's two-stage workflow: Plan mode analyzes the task and proposes an approach without making changes; Act mode executes the approved plan, applying file edits and running commands with approval required at each step. This human-in-the-loop design is Cline's primary safety mechanism, contrasted with more autonomous agents like Windsurf's Cascade.
Cursor's multi-file agent mode, capable of planning and executing changes across an entire codebase in a single session, including the newer background agent capability that runs on cloud VMs.
Windsurf's autonomous agent, notable for executing multi-step tasks with fewer pauses for approval than comparable tools — a deliberate design choice that trades user control for execution speed. See Cursor vs Windsurf for how this compares to Composer's approach.
An AI agent that runs on a remote cloud VM rather than your local machine, allowing it to continue working asynchronously — including while you are away from your computer — and to be triggered remotely (e.g., via Slack or GitHub). Cursor introduced this capability in February 2026, running up to 8 agents in parallel per codebase.
Claude Code's multi-agent orchestration feature, allowing multiple sub-agents to work in parallel on different parts of a complex task — for example, one agent handling a refactor while another writes corresponding tests simultaneously.
A saved snapshot of a workspace's state during an agent session, allowing the entire session — not just the most recent change — to be rolled back if the agent's work goes off-track. Cline's checkpoint system is a session-level safety net, distinct from git-based rollback (which Aider relies on instead).
An AI model at the current state of the art in capability — currently Claude Opus, GPT-5, and Gemini's top-tier models. Contrasted with smaller, faster models (Claude Haiku, GPT-4o mini) used for lower-stakes tasks like autocomplete, where speed matters more than maximum capability.
In Continue.dev's config.yaml, the specific function a configured model performs: chat, autocomplete, edit, agent, embed, or rerank. Assigning different models to different roles (a fast local model for autocomplete, a frontier model for agent tasks) is a core Continue.dev configuration pattern — see the Continue.dev Rules guide.
An AI model that runs entirely on your own hardware (typically via Ollama or LM Studio) rather than through a cloud API. Local models eliminate ongoing API costs and keep code from ever leaving your machine, at the cost of lower capability than frontier cloud models. Supported by Cline, Continue.dev, Aider, and Zed.
A widely cited benchmark that measures an AI model's ability to resolve real GitHub issues by generating a correct code patch, verified against the project's actual test suite. "SWE-bench Verified" is a human-validated subset. Scores are frequently cited in tool marketing (e.g., Claude Code's 87.6% with Opus 4.7) but reflect model capability on isolated issues, not full agent workflow performance — which is why we built our own task-based benchmark alongside it.
An open standard that lets AI agents connect to external tools and data sources — databases, deployment systems, documentation, web search — through a standardized interface. Cline has the most developed MCP ecosystem among the tools covered on this site, including a marketplace of community-built servers. Cursor and Continue.dev also support MCP, with varying levels of tooling around it. For a full practical walkthrough, see How to Set Up MCP Servers for AI Coding Agents.
An individual integration built on the MCP standard — for example, a server that gives an agent read access to a PostgreSQL database, or one that lets an agent search GitHub issues. Servers are typically installed and configured per-project or per-tool. See the MCP Setup Guide for configuration examples across Cline, Cursor, Continue.dev, Claude Code, and Windsurf.
A protocol, notably used by Zed, that allows an editor to connect to external AI agents (such as Claude Code) natively, rather than requiring the agent to be built into the editor itself. This is the mechanism behind combinations like "Zed + Claude Code" as a paired setup.
The two dominant pricing architectures in this category. Subscription tools (Cursor, Windsurf, GitHub Copilot, Claude Code) charge a flat or tiered monthly fee that includes model access. BYOK tools (Cline, Aider, Continue.dev) are free themselves, with the person providing their own API key and paying the model provider directly. See AI Coding Tools Pricing 2026 for a full breakdown of both models across every major tool.
A single repository containing multiple distinct packages or applications, typically managed with tooling like Nx, Turborepo, or pnpm/yarn workspaces. Monorepos pose specific challenges for AI tools around cross-package type resolution and context scoping — see Best AI IDE for TypeScript Monorepo for a full comparison of how tools handle this.
The mechanism (workspace:* in npm/pnpm, project references in Nx) by which packages within a monorepo reference each other locally rather than through the public npm registry. AI tools need to resolve these correctly to understand that an import like @myorg/types points to local code, not an external dependency.
A rule (often enforced by tooling like Nx) that restricts which packages within a monorepo are allowed to import from which other packages — preventing, for example, a shared library from depending on a specific application. AI agents that are unaware of module boundaries can suggest changes that technically compile but violate the intended architecture.
GitHub Copilot's project-level configuration file (located at .github/copilot-instructions.md), used to give Copilot standing instructions about a project's conventions. Does not transfer automatically when switching to another tool — see How to Migrate from GitHub Copilot to Cursor for how to convert this content to another tool's format.
A vendor's contractual commitment to defend and cover legal costs if AI-generated code triggers a copyright or intellectual property claim. Available on GitHub Copilot Business and Enterprise plans; less standardized across other tools. A common item on enterprise security checklists — see AI Coding Tools for Teams: Enterprise Guide.
Single Sign-On and Security Assertion Markup Language — the standard mechanisms by which organizations manage centralized identity and access across tools, including AI coding assistants. A near-universal requirement for enterprise procurement; availability and pricing for SSO varies significantly by tool (included in GitHub Copilot Business at $19/user, a $10/user add-on for Windsurf Teams).
System for Cross-domain Identity Management — a standard protocol for automatically provisioning and deprovisioning user accounts across tools when someone joins or leaves an organization. Relevant for enterprise rollouts where manual account management does not scale.