Cursor

Cursor is a coding agent product: a desktop editor, a terminal CLI, and cloud agents that work in their own virtual machines. Plans run from $0 to $200 a month, with third-party models like Claude and GPT billed at API list price on Premium and above. Checked 31 October 2026.

At a glance

What it is

The Cursor docs call it "a coding agent for building ambitious software." It began as an editor with AI in it. The current docs describe three working surfaces that share the same agent ideas:

• The desktop app. An editor with Agent and Chat, Plan mode, debug mode and a browser tool.
• The CLI. The same agent from a terminal, interactive or scripted.
• Cloud agents. Agents that run in isolated virtual machines with a full development environment, so they can build, test and interact with the software they changed.

Cloud agents are the newer idea. The docs say you can run as many in parallel as you like and they "do not require your local machine to be connected to the internet." You start them from the iOS app, Cursor Web at cursor.com/agents, the desktop app, Slack, a GitHub or Bitbucket comment with @cursor, Linear, or an API.

One more fact changes how you read the rest of this page. Cursor is now sold as part of the X Subscription, one subscription covering Grok, Cursor and X. The pricing page says subscriptions are only sold directly through X Subscription Corp and that no resellers are authorized. Our Cursor vs Claude Code page notes a company-ownership headline on Cursor's homepage; we leave the corporate story to Cursor and treat the product facts below as the point.

Plans and prices

From cursor.com/pricing and the models-and-pricing docs:

Annual billing is $80 a year for Plus, $300 for Premium, $1,000 for Super and $2,000 for Ultra. Prices exclude tax.

How to read it:

• Plus gets Agent with Cursor's own models at fixed settings, plus Grok Chat and related apps. It has no Auto, Cloud Agents, Automations, Bugbot or third-party models.
• Premium is where the product opens up: Cloud Agents and Automations, code hosting on Origin, Bugbot code review, third-party models and Grok Bot. The docs call it the plan for new Agent users.
• Super suits daily users, Ultra suits people running several agents or automations.
• Usage is one pool, shared across the editor agent, CLI, Cloud Agents and the Grok apps, and it refreshes weekly.
• People on older plans (Pro, Pro+, Ultra, Start, Teams Standard or Premium) are on "legacy Cursor plans" and keep them as they are, with two monthly usage pools and on-demand usage.

Our earlier duo page flagged that Cursor's pricing mixed several products. It still does, so the table above is about the Cursor-side rows only.

Install

Desktop app

Download from cursor.com/downloads, sign in, pick a folder and start with a small task. Requirements from the quickstart: macOS 12 (Monterey) or later on Apple Silicon or Intel, Windows 10 or later, and on Linux an apt repository for Debian and Ubuntu, a yum repository for RHEL and Fedora, or an AppImage.

# Debian/Ubuntu: after adding Cursor's GPG key and repository
sudo apt update
sudo apt install cursor

CLI

# macOS, Linux, WSL
curl https://cursor.com/install -fsS | bash

# Windows PowerShell
irm 'https://cursor.com/install?win32=true' | iex

agent --version
agent "refactor the auth module to use JWT tokens"

If agent is not found, add ~/.local/bin to your PATH. The CLI supports an interactive session and print mode for scripts and CI pipelines.

The CLI command is agent, not cursor. That trips up newcomers, so check it first when something seems missing.

Models

The docs list models from Anthropic, OpenAI, Google, Moonshot, Meta, DeepSeek, Z.ai and Cursor itself. Model pages exist for Claude Sonnet 5.5, Opus 5.5, Fable 5.1 and Haiku 5.5, Gemini 4 Argon and 3.8 Flash, GPT-5.6 Sol, Terra and Luna, Grok 4.7, 4.6 and 4.5, and Composer 2.5. On Teams and Enterprise, Cursor Router can pick among Cursor and third-party models for each request based on your optimization mode.

Two rules matter more than the list:

1. Cursor models draw on your plan's weekly pool, and third-party models are billed at API list price on top.
2. Some models need data-retention approval when Privacy Mode is on. The docs say Claude Fable 5, for example, needs approval for Enterprise, Teams and individual users with Privacy Mode, because Anthropic stores its input and output for harm prevention without using it for training.

Rules, skills, hooks and MCP

Cursor has four ways to give the agent standing instructions:

Skills follow the open Agent Skills standard. Cursor discovers them at startup, the agent decides when they apply, and you can call one by typing / in chat. Built-in skills include /automate (create Automations triggered by schedules, Slack messages or GitHub events), /autopilot (watch a pull request and handle feedback, conflicts and failing checks), /canvas and /create-hook.

Hooks run your scripts around the agent loop, defined in hooks.json. The docs list uses like running formatters after edits, scanning for secrets, and gating risky operations such as SQL writes. Cursor can also load hooks from third-party tools like Claude Code.

MCP supports three transports: stdio (local, Cursor manages it), SSE and streamable HTTP (local or remote, with OAuth). It supports tools, prompts, resources, roots, elicitation and MCP Apps. Configure servers on the Customize page or in mcp.json. Good first servers: Playwright MCP, GitHub MCP and Context7. See also the MCP server list.

Cloud agents in practice

A cloud agent is an agent that runs on Cursor's side, in a virtual machine that looks like your laptop: cloned repositories, installed dependencies, secrets, startup commands and network access. The docs give four reasons to use one.

• You can run many in parallel.
• Your machine does not need to stay online.
• The agent can build, test and interact with the software it changed, including controlling a desktop and browser.
• It supports MCP servers, so it can reach databases, APIs and other tools.

A cloud agent can also work across several repositories at once. If a change spans a frontend, a backend and a shared library, one agent can inspect the full workspace and open pull requests in each repo it changes.

Before anyone can start one, an account admin has to connect source control. The docs list GitHub (cloud and Enterprise Server), GitLab (cloud and self-hosted) and Bitbucket Cloud. The cloud-agent section of the docs also covers self-hosted runtimes, private connectivity and network security for teams that cannot send code to Cursor's machines.

Cloud agents begin at Premium. They are the biggest practical difference between the $8 Plus plan and everything above it.

Safety and privacy

Run modes control how much the agent can do without asking. In Settings, Agents, Approvals & Execution you choose between:

The docs call Auto-review the safest useful setup, and also say plainly that "Auto-review is not a security boundary" because the classifier can make mistakes.

Privacy Mode. The security page says: "When enabled, we will not train on your data," and that Privacy Mode adds technical controls and contractual requirements with model providers. It is available to anyone, free or paid, and a team or enterprise admin can enforce it. Cursor holds a SOC 2 Type II attestation, with reports on request.

What third-party models add to the bill

The plan price is only part of the cost on Premium and above, because third-party models are billed at API list price from prepaid credit. You can estimate that part from the model's own price. Take Claude Sonnet 5.5 at $2 per million input tokens and $10 per million output tokens, which is Anthropic's published rate. The token volumes below are assumptions to show the arithmetic. Your own usage will differ, and the Cursor dashboard shows the real figures.

Two lessons come out of the table. At light use the plan dominates, so pick the plan for its features, such as Cloud Agents and Bugbot. At heavy use the model spend dominates, and the Super or Ultra upgrade only helps if the extra pool covers Cursor's own models you would use instead. If most of your work runs on Claude or GPT, the plan tier barely changes the bill, and the cost looks like an API bill with an editor attached. The reverse is true if you stay on Cursor's own models, where the weekly pool is what you are buying.

On Teams and Enterprise, eligible models also carry the Cursor Token Rate, which the pricing page puts at $0.25 per million tokens on top of the API price. Add that to any estimate for a team. Agent loops resend context on every step, so cost scales with how long a session runs, not with how many prompts you typed.

A first-week plan

The Free plan is enough to learn the shape of the product. A week is a fair trial.

Write down what you spent and what annoyed you. A week of notes beats any feature table, including ours.

Rules that keep a cloud agent safe

A cloud agent works while you are away, so it needs tighter limits than an interactive session. The docs describe the environment a cloud agent gets: cloned repositories, installed dependencies, secrets, startup commands and network access. Each of those is a decision for you.

• Secrets. Give the agent only the secrets its task needs, and prefer read-only credentials for databases and APIs. A secret in the environment can appear in logs or in a pull request description.
• Network. Limit outbound access to the hosts the build needs. The docs cover network security for teams, so use it before enabling cloud agents on a private repository.
• Branches. Let agents open pull requests and never push to the main branch. Protect that branch with required reviews.
• Scope. One task per agent, with a written definition of done. A vague brief produces a large, hard-to-review diff.
• Review. Read the pull request as you would a colleague's. Run the tests locally if the change touches anything sensitive.
• Cost. Parallel agents all draw on the same pool. Ten agents on ten tasks is ten times the usage, so start with two.

Common problems

Who it suits

• Developers who want the editor as the home screen, with the agent next to the code.
• People who want to run several agents in the cloud and review pull requests.
• Teams that want admin controls: SSO, spend controls, audit logs and model and MCP access rules on Enterprise.
• Anyone who wants a CLI and an editor on one usage pool.

It suits less well if you want an open-source agent, a flat tool price without model-usage charges, or only a terminal.

Moving to it from another agent

1. Install the desktop app and open your repository.
2. If you have an AGENTS.md, you are done with instructions; otherwise move the content into .cursor/rules or AGENTS.md.
3. Recreate MCP servers in mcp.json.
4. Move reusable prompts into skills, since the standard is shared with other agents.
5. If you use Claude Code, you can keep it: the Claude Code docs offer an "Install for Cursor" link for the same extension, so both can run in one window.

Alternatives and duos

Our take

This comes from Cursor's pricing page and docs, not from a test.

The product is broader than "an editor with AI." An editor, a CLI, cloud agents, code review and an automation layer share one weekly usage pool, which is a coherent design. The docs are also candid about limits, such as the Auto-review warning.

The thing to examine before paying is the bundle. The pricing page now spans Cursor, Grok and X benefits, with Cursor's own models in the base pool and third-party models billed separately. If you mostly want Claude or GPT models, work out the API-rate cost on top of the plan price. If you mainly use Cursor's own models, the pool is the cost.

Try it on the Free plan first, switch Privacy Mode on, and spend a week on a real task before choosing a tier.

Checklist

• Choose a plan and note which models it includes
• Turn on Privacy Mode and ask your admin to enforce it for teams
• Set Run Mode to Auto-review
• Add rules to .cursor/rules or an AGENTS.md
• Add one MCP server
• Install the CLI and check agent --version