Install Cursor and open a project folder
Cursor is a code editor, an app for reading and changing the files that make up a program, with an AI agent built in. An agent is an AI that plans a task and then acts on your files to finish it. In Cursor it is simply called Agent. You type a request in plain words, and Agent finds the files involved, edits them and can run commands to check the result.
You need a Cursor account, which you can create at cursor.com, and a computer on Windows 10 or later, macOS 12 Monterey or later, or Linux. Macs with Apple Silicon and Intel chips are both supported. Setup takes a few minutes.
Agents can also run from a web page, cursor.com/agents, through a feature called Cloud Agents. This lesson stays with the desktop app, where you see every file change on your own machine.
- Download. Go to cursor.com/download and click the button for your operating system. Windows gets an .exe installer and macOS a .dmg file.
- Install. On a Mac, drag Cursor into the Applications folder. On Windows, run the installer and follow the prompts. On Linux, install it with apt or dnf, or run the portable AppImage file.
- Sign in. Open Cursor and sign in with your Cursor account when it asks.
- Open a folder. Choose File > Open Folder and pick the folder that holds your project. For a first try, make a new empty folder and open that.
- Open Agent. Press Ctrl+I on Windows or Linux, or Cmd+I on a Mac, and the Agent panel opens beside your files.
On Linux, Cursor prefers its apt and dnf packages to the AppImage. The packages add a desktop icon, automatic updates and command-line tools.
Ask Agent to explain, then to change one thing
A codebase is every file of a project taken together. On a project you did not write, start with a request that only reads. Cursor's quickstart opens with the two prompts below, copied as the guide gives them. One maps the project. The other asks for ideas and makes Agent wait for your pick before it edits anything.
Good first jobs are small and low risk, such as fixing the wording of a heading or a button that sits out of line. When you already know the change, ask for it directly and describe the result you want. Cursor's help center uses this request as its example: "Add a dark mode toggle to the settings page."
Explain this codebase. Point me to the main entry points, key modules, and anything I should read before making changes.
Suggest three small, safe improvements in this codebase. Explain the tradeoffs and wait for me to choose one.- Point at files. Type @ in the chat box and then a file or folder name to tell Agent where to look. It also searches the codebase by itself.
- Queue the next request. Type your next instruction while Agent is busy and press Enter. It runs once the current task ends, and you can drag queued messages into a new order.
- Stop it. Click Stop to halt Agent in the middle of a task, then redirect it or try another way.
- Answer its questions. Agent may ask you something during a task. It keeps reading and editing while it waits, and uses your answer as soon as it arrives.
- Send selected lines. Select some code and press Ctrl+L, or Cmd+L on a Mac, and Agent opens with those lines already in the input.
- Ask on the side. Type /side in the chat box, followed by your question, to open a side chat. It looks into a tangent without interrupting the main conversation.
Plan Mode
Agent researches the codebase, asks questions and writes a plan you can edit before it builds.
Source: Cursor blog: Plan Mode (read November 5, 2026)
/goal and steering
/goal gives Agent a long-lived objective to work toward, and a message sent mid-task reaches it at its next tool call.
Source: Cursor changelog (read November 5, 2026)
Cloud Agents without a repository
A Cloud Agent can start a project from scratch, with no GitHub connection needed first.
Source: Cursor changelog (read November 5, 2026)
Remote control for local agents
The Cursor iOS app shows the agents running on your computer and lets you reply to them.
Source: Cursor changelog (read November 5, 2026)
Review every edit and roll back with checkpoints
Agent applies its edits while it works, and they are saved to disk straight away. A diff is the view that marks each line removed and each line added, file by file. Read it when Agent finishes, and reject any change you did not ask for.
Then ask Agent to run the checks your project already uses. Checks are commands that test the code for you: unit tests, a type checker, a linter that flags style errors, or a local build. A new empty project has none of these yet. Open the page in a browser and click through it yourself, the way you checked the Lovable app.
Agent can look at the page too. Its browser tool opens the app, clicks through it and takes screenshots to confirm that a visual change landed. You can ask it to use that tool on whatever it just changed.
A checkpoint is a snapshot of your files that Agent saves before a significant change. If Agent takes a wrong turn, hover over an earlier message and click Restore Checkpoint. Every file Agent changed after that point goes back. Restoring reverts files only, so the messages stay in the chat.
Checkpoints are stored on your computer, separate from Git. Git is a version control tool: it records each saved state of a project, called a commit, so you can return to any of them later. Use checkpoints to undo Agent's work within a session, and Git for history you want to keep.
For a second opinion, type /agent-review in the Agent input. It runs a separate code review of your local changes. A Quick review is fast and suits small edits. A Deep review is slower and costs more, and fits security-sensitive code or a large rewrite.
Auto-reload, a setting that restarts a running app whenever a file changes, reverses this order. With it on, Agent's edits may run before you have read them.
Pick a mode and decide what runs without asking
Press Shift+Tab in the Agent input to cycle through the four modes, or pick one from the dropdown. Each mode starts a fresh context, the working memory of a conversation. Start a new chat when you change tasks, too.
Plan earns its place when a change spans several files or needs research. Instead of writing code at once, Agent searches the codebase for the files involved, asks you questions, writes a detailed plan and waits for your approval before it builds.
Some actions never ask. Agent reads files and searches code without approval, and it edits files in your folder without asking, except configuration files such as workspace settings. To keep a file out of its reach, list it in a .cursorignore file in the project. Its web access is narrow as well. With default settings, Agent's tools reach only GitHub, links you hand it and web search providers.
Terminal commands work differently. A terminal command is a typed instruction to the computer, such as installing a package or starting the app, and by default each one waits for your approval. Tools you connect through MCP, a standard way to plug outside services into an agent, ask too: once for the connection, then for every call.
Run Modes, under Settings > Agents > Approvals & Execution, change how many prompts you see. Allowlist runs the actions you have listed without asking. Auto-review runs listed calls at once and runs other shell commands in a sandbox, a fenced-off space that cannot freely reach protected files or the network. Anything left goes to a classifier, a small AI model that judges each call. Run Everything runs every call with no prompt at all.
Cursor's Run Modes page is blunt about the middle option: "Auto-review is not a security boundary." The classifier can allow a call you would have blocked. While you are learning, keep the prompts and read each command before you click.
One more default matters when you download a stranger's project. Workspace trust, a setting that opens unknown folders in a restricted mode, is off by default in Cursor. Cursor's security page suggests a basic text editor for a repository you do not trust.
| Mode | Best for | Edits files? |
|---|---|---|
| Agent | Building features, fixing bugs, reorganizing code | Yes |
| Ask | Understanding code and how the parts fit | No, read-only |
| Plan | Bigger features: it researches, asks questions and writes a plan first | After you approve the plan |
| Debug | Tricky bugs that need evidence from the running app | Yes |
Write rules so you stop repeating yourself
A rule is a standing instruction Agent follows in every chat on a project, such as which language new files use. Without one, you type the same correction again and again. Start small, and add a rule when Agent makes the same mistake more than once.
To make one, open the command palette with Ctrl+Shift+P on Windows or Linux, or Cmd+Shift+P on a Mac, and choose New Cursor Rule. Name it and write the instructions in Markdown, a plain text format where a dash starts a list item. Cursor saves the file in .cursor/rules/ inside the project, so it travels with the code.
A type setting decides when the rule applies: Always Apply, Apply Intelligently, Apply to Specific Files or Apply Manually. A plain AGENTS.md file in the project's top folder also works, and Cursor reads it automatically. The three lines below are a starting rule written for this lesson.
Rule types, file patterns and moving an old .cursorrules file are all covered in Mindket's guide to Cursor rules, which goes much further than this section. Keep each rule under 500 lines, and split a long one into focused files.
Rules reach Agent only. Tab completion, the code suggestions that appear as you type, ignores them. Cursor also reads a CLAUDE.md file in the top folder the same way as AGENTS.md, so instructions written for another agent carry over.
- Explain each change in one plain sentence before you make it
- Keep every page readable on a phone
- Ask before adding a new packageHow Cursor's co-founder and Lee Robinson run Agent
Michael Truell, a co-founder of Cursor, posted the team's tips on X on January 20, 2026, quoting Cursor's guide to working with agents. Start with a plan in Plan mode, which Shift+Tab reaches. Let Agent search the code on its own instead of tagging every file with @. Use tests as the feedback loop. When a run goes sideways, revert, tighten the plan and run it again.
In this lesson's terms, that last tip is a short routine. Click Restore Checkpoint on the message before the bad run, edit the plan where it was vague, then send it again. Add a rule once the same mistake comes back, as his list also suggests.
Lee Robinson, who was a vice-president at Cursor and now works on product at SpaceX by his X bio, described his own habits on X on December 31, 2025. He starts every new feature with a plan in Plan Mode. He reviews the code three times: while Agent writes it, with an agent review in the editor before he pushes, and with Bugbot, Cursor's AI reviewer for pull requests. The middle pass is the /agent-review command from this lesson. His own 2026 course below teaches the same loop in 30 minutes: planning features, fixing bugs, and reviewing and testing code.
When it goes sideways: revert → tighten the plan → rerun
A screenshot of X's own embed of the post, taken on 5 November 2026, with the time in UTC. X cuts the post short with Show more. Select it to open the post on X.
Next: GitHub Copilot agent mode
Before you move on, run one round trip in Cursor. Open an empty folder and ask Agent for a one-page site from the brief you wrote in the first lesson. Read the diff and open the page. Then ask for one fix, restore the checkpoint from before it and watch the file change back.
Copilot brings the same habits to VS Code, Microsoft's code editor. The GitHub Copilot lesson sets up agent mode there. It also covers an agent that works on GitHub's own servers and hands its changes back for review.
Check yourself
Agent changed three files and you want all of them as they were before your last request. What do you click?
Show answer
C. Restore Checkpoint Restoring a checkpoint returns every file Agent changed after that message, and the chat stays as it is.
Which mode answers questions about the code without editing a single file?
Show answer
A. Ask Ask mode is read-only, so it explains and explores but never changes files.
With the default settings, which action waits for your approval?
Show answer
D. Running a terminal command Reads, searches and most file edits go ahead, while terminal commands ask first unless you change Run Modes.
Where does a rule made with New Cursor Rule get saved?
Show answer
B. In .cursor/rules/ inside the project Project rules are files in .cursor/rules/, so they live with the code and can go into Git.
Frequently asked questions
Is Cursor free?
You can start without paying. Creating an account costs nothing, and the Free plan comes with limited usage. Heavier use of Agent needs one of the paid plans on Cursor's pricing page.
Is Cursor the same as VS Code?
Close, but no. It uses VS Code's default keyboard shortcuts and adds its own for the AI features, such as Ctrl+L to send selected lines to Agent. VS Code itself gets an agent through GitHub Copilot instead.
What is the difference between Agent and Ask in Cursor?
Ask only answers. You can put questions to it such as where the database connection is configured, and it leaves every file untouched. Agent answers too, then goes on to edit and run commands.
What happened to .cursorrules?
The single .cursorrules file at a project's root is the legacy format, due to be retired. Move its text into a new rule, choose Always Apply to match the old behavior, then delete the old file.
Sources
- Cursor docs: Quickstart https://cursor.com/docs/get-started/quickstart
- Cursor help: Download and install Cursor https://cursor.com/help/getting-started/install
- Cursor help: Your first project https://cursor.com/help/getting-started/first-project
- Cursor help: Agent mode https://cursor.com/help/ai-features/agent
- Cursor help: Ask mode https://cursor.com/help/ai-features/ask-mode
- Cursor docs: Cursor Agent https://cursor.com/docs/agent/overview
- Cursor docs: Agent Security https://cursor.com/docs/agent/security
- Cursor docs: Run Modes https://cursor.com/docs/agent/security/run-modes
- Cursor docs: Agent Review https://cursor.com/docs/agent/agent-review
- Cursor help: Rules https://cursor.com/help/customization/rules
- Cursor help: Keyboard shortcuts https://cursor.com/help/customization/keyboard-shortcuts
- Cursor help: Pricing https://cursor.com/help/account-and-billing/pricing
- Cursor blog: Plan Mode https://cursor.com/blog/plan-mode
- Cursor blog: Best practices for coding with agents https://cursor.com/blog/agent-best-practices
- Cursor changelog https://cursor.com/changelog
- Cursor changelog, page 2 https://cursor.com/changelog/page/2
- Wikipedia: Cursor (company) https://en.wikipedia.org/wiki/Cursor_(company)
- Michael Truell on X, January 20, 2026 https://x.com/mntruell/status/2013636888242835810
- Lee Robinson on X, December 31, 2025 https://x.com/leerob/status/2006413745380114833
- leerob on YouTube: Cursor: coding agents tutorial (2026) https://www.youtube.com/watch?v=kF2WQgk1LtY