boxes.dev vs Cursor

At a glance

As of September 2026.

Whose agent does the work

Cursor runs its own agent with a choice of models, and boxes.dev runs Claude Code and Codex themselves.

Cursor's agent works in the editor, the CLI, and cloud agents. You can choose Cursor's own models or models from Anthropic, OpenAI, Google, and others. Cloud agents bill at the selected model's API price against the usage pools in your Cursor plan, then on-demand, and Cursor publishes no separate charge for the VM.

boxes.dev runs the Claude Code and Codex CLIs on your devboxes, signed in with your Claude or ChatGPT plan or an API key, so their commands, skills, and settings work as they do on your computer.

How the cloud environment is built

Cursor prepares cloud environments from your repositories and an install command, and boxes.dev copies a machine where your app already runs.

Cursor's agent can set up a cloud environment for you, or you can supply a Dockerfile. Cursor clones each repository, runs the install command, and saves the disk as a snapshot, then rebuilds on a schedule to stay close to your default branch. Each new agent starts from the latest snapshot and runs start commands to bring up Docker, databases, and app processes.

boxes.dev first sets up your full development environment on a cloud machine called the Template box, with the tools, databases, services, and data your app needs, so an agent can run the app and test its changes end to end. A setup agent can port the environment from your laptop, including the files, local data, agent settings, and skills you choose to upload, or build it from a GitHub repository. boxes.dev then saves the Template box as a snapshot, and every task runs on its own devbox, a full, isolated copy of that machine. A maintenance script can keep the Template box current as your code changes, running on a schedule or when files such as a lockfile change, and each successful run saves a new snapshot.

You'd never let an engineer ship code they hadn't run, and a full environment holds the agent to the same standard. On a boxes.dev devbox, the agent runs your whole app, with its databases and services, and tests its changes in the browser as it builds, and it reproduces bugs in the real app to find the root cause instead of guessing from the code, so it writes better code and makes fewer mistakes. It can also take screenshots and record video for you to review, so you can hand it most repetitive manual testing.

When the machine goes idle

A boxes.dev devbox keeps its running programs through sleep, and a Cursor cloud agent keeps its disk.

After a run goes idle, Cursor hibernates the VM and later deletes it. A snapshot of the disk is kept for 90 days of inactivity, so the agent can resume without cloning again.

A boxes.dev devbox sleeps when nobody is using it, and an agent that is still working keeps it awake. Sleep keeps the machine's files and memory, so dev servers, databases, and terminal sessions resume where they left off when it wakes. The devbox stays yours until you delete it.

Reaching the running app

Both let you use the app running in the cloud, and boxes.dev also gives each devbox public links and your project a stable URL.

In Cursor's Agents Window, you can take control of a cloud agent's desktop to use the app yourself, or forward the VM's ports to your computer and open the app in Cursor's browser, where Design Mode lets you point the agent at elements. To publish a live URL, Cursor deploys the repository to Vercel.

Each boxes.dev devbox runs a browser that you and the agent share, so you can watch the agent test the app, click through it yourself, or comment on page elements to show the agent what to change. Ports forwards the app to the same localhost port on your laptop, a public link shares a devbox's running app with anyone, and a Project URL gives your project one stable public address that you can point at any devbox, which suits OAuth callbacks and webhooks.

Working with your team

Both bring agents into Slack and share agent work with teammates, and both control whether a teammate can direct an agent that holds someone else's access: a team admin decides in Cursor, and you decide for your own devboxes in boxes.dev.

Mentioning @cursor in Slack starts a cloud agent, and later mentions in a thread with an agent add follow-up instructions. Teammates on your Cursor team who can access the repository can open an agent's link, read-only unless a team admin allows team follow-ups. Cursor's docs warn that a teammate's follow-up runs with the creator's secrets and credentials and could escalate that teammate's access, so admins can allow follow-ups on every agent, only on agents that Enterprise service accounts start, or on none.

Each boxes.dev devbox belongs to one person and works with their access, such as their GitHub login and cloud credentials. Anyone who can prompt its agent could act with that access, so in the app, teammates get read-only access to your work. You can share a thread as a link, and a team admin can let teammates read every thread that isn't marked private and browse each other's projects in a Team workbench. Teammates can't send prompts, open your terminal, or control your devbox.

In Slack, mentioning @boxes.dev in a thread starts Claude Code or Codex on a new devbox, and later mentions in that thread send follow-ups to the same conversation. By default, only you can start work in your project from Slack or follow up on it. If you allow teammates, they can also start work there with @boxes.dev @YourName and send follow-ups, and their requests run with your access and usage. boxes.dev sends you a DM with links to each task a teammate starts, and commits can link to the thread that produced them.

What comes with every boxes.dev devbox

The boxes.dev desktop app opens the same set of tools for whichever devbox you select:

• Terminal, with tabs that keep running after you close the app.
• Files, a file browser and editor with search across the workspace.
• Git/Review, for branch history and diffs, where comments on changed lines go to the agent together as one message.
• Browser, the browser on the devbox that you and the agent share.
• Ports, for the forwarding, .localhost addresses, and public links described above.
• Activity, which shows the processes using the machine's CPU and memory.

Where Cursor shines

• The editor. Tab completions, inline edits, and the agent all work in one AI code editor.
• Model choice. One plan covers Cursor's own models and models from Anthropic, OpenAI, Google, and others.
• Pull request agents. Bugbot reviews pull requests for bugs, Security Agents look for vulnerabilities, and PR Routing & Approval assigns reviewers and can approve low-risk changes.
• Event-driven automations and an SDK. Automations can start on pull request and push events, and TypeScript and Python SDKs run Cursor's agent from your own code.
• A desktop for every agent. Each cloud agent controls a full desktop with a mouse and keyboard and can hand control to you.

Pricing

As of September 2026.

A box-hour is one default devbox awake for an hour, and sleeping devboxes don't use any. On Starter, 40 box-hours works out to about two hours of devbox time each working day, and Pro's 250 to about twelve, spread across as many as ten devboxes at once. Your Claude or ChatGPT plan covers the agents' usage. See Plans, seats, and box-hours.

Your Claude or ChatGPT plan is also heavily discounted. By our estimate, when fully utilized, a Claude Max plan is at least 40x cheaper than the same usage at API prices, and the $200 ChatGPT Pro plan is over 30x cheaper. So on boxes.dev, you'll pay far less for the same tokens.

Cloud agents draw on the same usage pool as the editor, so heavy agent use adds on-demand spending at API prices. On Teams, third-party models also carry a Cursor rate of $0.25 per million tokens. See Cursor's models and pricing.

Using both

You can keep Cursor as your editor and use boxes.dev when you want Claude Code or Codex on a full cloud machine. Both can work on the same repositories and pull requests.

Frequently asked questions

Does Cursor run Claude Code?

No. Cursor offers Claude models inside its own agent, and Claude Code is Anthropic's separate agent. boxes.dev runs Claude Code itself, signed in with your Claude plan or an API key.

Can Cursor's cloud agents run my database and services?

Yes. Start commands can bring up Docker, databases, and app processes each time an agent starts. On boxes.dev, each devbox is a copy of your Template box, so its databases and seeded data are already there.

Can I open a devbox in my own editor?

Yes. dvb ide opens a devbox in VS Code or a JetBrains IDE over SSH. See Install and use the CLI.

Sources

• Cursor: cloud agents, cloud agent setup, Builds, capabilities, security, settings, service accounts, automations, Slack, CLI, subagents, Cursor for iOS, models and pricing, and plans.
• boxes.dev: How boxes.dev works, Update your Template box and snapshots, Startup, teardown, and maintenance, Sleep, wake, and recover a devbox, Review a page with your agent, Files and transfers, Preview your app, Project URLs, Use the agent's terminal view, Connect Codex and Claude, Shared project storage, Organize and share threads, Team access and roles, Commit, push, and open a pull request, Install and use the CLI, Work from Slack, Work from Linear, Use boxes.dev on your phone, and Plans, seats, and box-hours.