Claude Code runs in the terminal, in VS Code and JetBrains, in a desktop app, and on the web. Cursor’s Agent sits in the editor side pane (Cmd+I / Ctrl+I), and Cursor also has a command line you start with agent. Both cover command-line and editor work. The choice that matters is the window you already work in — the terminal, or an editor that shows inline diffs next to the code — plus the instruction files you keep in the project, and the habit of planning first and rewinding when a turn goes wrong.
Review
Claude Code shows the diff, the before-and-after of changed files, in the terminal, in the VS Code extension, and in the desktop app. In Plan Mode, Claude must show a written plan and wait for your OK before it edits files. The terminal and editor views already show the file changes for you to review. You can also create a named skill for a repeated workflow, then run it using the name you chose.
Cursor Agent applies edits in the editor. While it works you can queue the next instruction, or send a follow-up that takes effect when the current step finishes. The command line uses the same three modes as the editor: Agent to make changes, Plan to design the approach first, and Ask to explore read-only without writing files.
If you want a written plan before any write, Claude Code’s Plan Mode and Cursor’s Plan mode are the closest match.
Checkpoints
Both keep session-level snapshots so you can undo agent edits without reaching for git.
Claude Code creates a checkpoint before each prompt and tracks edits Claude made by writing files itself. /rewind, or Esc twice on an empty prompt, opens a menu to restore code, conversation, or both. Snapshots do not include changes made only through terminal commands (bash), and they are not a substitute for commits.
Cursor Agent creates checkpoints before significant changes and shows them on the chat timeline. You can preview files at a checkpoint and restore them. Cursor also stores those snapshots locally, separate from git.
Use a checkpoint to undo the file edits from a prompt when you want to try a different approach. Use git for anything you want to keep.
Instructions and context
Claude Code reads CLAUDE.md at the start of every session. Skills load when you call them. Put the stack, conventions, and review checklist in CLAUDE.md — things Claude should see every time, such as “this repo is TypeScript; run the review checklist before you stop.” Put a step-by-step workflow in SKILL.md and call it when you need it, such as a staging deploy or a local diff review.
Cursor Agent follows a system prompt plus project rules. The command line can run non-interactively in scripts: agent -p prints a result and exits. Both tools expect project knowledge somewhere the agent will see without you pasting it each time.
A small job with a clear stop is easier to judge in either tool than a long, open-ended session.
Try one small job
Stay in the window you already have open. Give both products the same small job. This is an editorial example, not a test we ran:
Add a due date to the task list. Include a test. Do not change login or the list layout. Stop when the test passes.Write that stop condition in the prompt. Then notice what you actually used — a plan you approved, a diff in the editor, a checkpoint you restored.
If that window is a terminal, start with Claude Code’s command line or Cursor’s agent. If it is an editor with inline diffs, start there. Switch only after that one job, once you know which review flow you prefer.