Surface - The environment where the work runs: the IDE, the Agents Window, the CLI, a Cloud Agent or the API. Each has its own context, permissions and security boundary.
Tab - A prediction of the next edit right in the editor: accept, reject or jump to the next place; it draws on recent edits, surrounding code and linter errors, and a suggestion can lead into another file.
Inline Edit - A pointed change to the selected fragment from a short instruction, without starting a full agentic loop.
Agent - A loop with tools: it reads files, edits code, runs commands and returns a diff; it stops according to the Run Mode and approvals.
Modes (Ask, Plan, Agent, Debug, Design) - A choice by degree of uncertainty: understand, design, change or find the cause of a failure. The mode sets what the agent is allowed to do.
Context - Everything the model actually sees: @-mentions of files and folders, terminal output, the git diff, past chats, browser pages and the repository index.
Indexing and .cursorignore - Building the repository search index and explicitly excluding paths from access and from the index - the boundary of what the agent can read at all.
Checkpoints - Snapshots taken before the agent's edits: they let you roll a change back inside the work; they do not replace git as the real history.
Run Modes - The level of autonomy for running commands: auto-review with a classifier check, an allowlist of permitted commands (an empty list means approving every command) and Run Everything, where calls go through automatically, though browser, file-deletion and external-file protection may still ask.
Local sandbox - Restricting commands with operating-system means: what the process may read, write and connect to over the network; it differs by platform.
permissions.json - Lists of terminal commands and MCP tools that run without approval, plus text guidance for the auto-review classifier; there are no denies in the file, file and network access lives in sandbox.json, and the CLI has its own permission system.
Rules and AGENTS.md - Durable instructions for the project and the user: repository rules, team rules and the AGENTS.md file as a portable contract.
Skills - A repeatable procedure as a product: a folder with a description, steps, scripts and assets, loaded by relevance or by an explicit call.
Subagents - Separate roles with their own context and tool set - exploration, shell, the browser or a role of your own for a task.
Hooks - Deterministic reactions to lifecycle events: allow, change or block an action and leave a trace for auditing.
MCP - Connecting external tools and data over the Model Context Protocol via stdio, HTTP or SSE, with its own authorization and tool policy.
Cloud Agent - A task executed in an isolated cloud environment with its own setup, secrets and network policy; the result comes back as a branch and a pull request.
Automations - Running an agent on a schedule or an event with defined tools, memories and stop conditions - that is already a production process, not a one-off run.
Bugbot and security agents - Automatic review of changes and a specialized hunt for vulnerabilities as separate layers on top of local review and the human on merge.