Cursor Cloud Agents, formerly Background Agents: what they cost

Cursor Cloud Agents are the half of Cursor that works while you are not looking. You hand one a task from the editor, from Linear, from a GitHub comment, or from Slack, it clones your repository into an isolated machine on Cursor's infrastructure, and it comes back with a branch and a pull request. They used to be called Background Agents; Cursor's docs now say so in one line and everything else on the page uses the new name. If you are deciding whether Cloud Agents are the right place to send the tickets nobody has started, the questions are how you start one, what it costs on each plan, and the handful of things it will not do.

Checked in early September 2026 against Cursor's own docs and pricing pages, linked throughout, and re-checked on October 16, 2026, after Cursor moved to The X Subscription. Cursor ships changes weekly; the pages are the source of truth.

Sinatra is our product and it competes with Cloud Agents on the Linear-to-pull-request loop, so this is a competitor writing about a competitor. We have tried to be plain about where Cursor is the better choice, which for many readers of this post it will be.

Cursor's Cloud Agents page opens with "Cloud Agents were formerly called Background Agents." The product underneath is the same idea: an agent that runs on Cursor's machines rather than yours, so it can keep working after you close the laptop, and so several can run at once without fighting over your checkout.

The name change matters for one practical reason. A lot of the writing online, including forum threads about pricing, uses the old name, and Cursor's own pricing docs have moved on. If you are searching for Cursor background agents, the current docs are under Cloud Agents.

Cursor's Cloud Agents page lists seven entry points: Cursor for iOS, Cursor Web at cursor.com/agents, the desktop app by selecting Cloud under the agent input, Slack with the @cursor command, a comment of @cursor on a GitHub pull request or issue (or a Bitbucket pull request), Linear, and an API. Its integration docs add two more: a mention of @Cursor in Microsoft Teams, and Jira, which is for Teams and Enterprise plans only.

The Linear path is the one a product manager will use. Per the Linear integration docs, you open the issue, click the assignee field, and select Cursor. You can also mention @Cursor in a comment to start a new agent or give an existing one more instructions. The agent shows real-time status in Linear and creates the pull request automatically when it finishes.

The GitHub path is a comment. Mention @cursor on a pull request or on an issue and an agent kicks off with that context. It is a mention rather than a label or an assignment, which is a small difference in a demo and a larger one when you want a non-engineer to be able to hand work over without writing a prompt.

The agent gets an isolated virtual machine with your repository cloned and, if you have set one up, your environment configured. Per the Cloud Agents page, it works on a separate branch, pushes the changes to your repository for handoff, and produces screenshots, videos, and logs to show what it changed and how it checked the work. On the Linear path, the Linear docs say agents show real-time status in the issue and create pull requests automatically when they finish.

Steering is mostly by mention. Reply @cursor on the pull request or @Cursor on the Linear issue and the agent picks up the instruction. There is now a second way: per the capabilities docs, an agent can subscribe to its own pull request, end its turn, and wake when review comments or CI results arrive, without you prompting it again. It only does that when the prompt asks it to, for example "open a PR and keep CI green," so a plain changes-requested review on an agent that wasn't told to watch still needs a mention. Automatic review of pull requests is a separate Cursor product, Bugbot, which reads diffs rather than writing them.

Models are "a curated selection," in the docs' words. You pick from the list Cursor offers for cloud work, which is not the same as the list you get in the editor.

Cursor is now sold as part of The X Subscription, one plan for Cursor, Grok, and X. Cloud Agents need Premium or higher. The pricing docs list what Plus leaves out: "No Auto, Cloud Agents, Automations, Bugbot, Origin, third-party models, or your own API keys," and Free is limited agent and chat use with no Cloud Agents either.

Plan	Price	Cloud Agents
Free	$0	No
Plus	$8/mo	No
Premium	$30/mo	Yes
Super	$100/mo	Yes
Ultra	$200/mo	Yes
Teams	$30, $100, or $200 a seat a month	Yes
Enterprise	Custom	Yes

Annual billing is $300 a year for Premium, $1,000 for Super, and $2,000 for Ultra. A Teams seat gets the same weekly usage as the individual plan at the same price. Customers already on the older Pro ($20), Pro+ ($60), Ultra ($200), or Teams Standard and Premium seats ($40 and $120) keep those plans; new customers start on The X Subscription.

The plan is the entry fee, and what the agent's work costs depends on the model. Cursor's own models, Grok and Composer, draw on a weekly usage pool that every paid plan includes. Cursor sizes it relative to Plus, so Premium is "4x Plus" and Ultra is "50x Plus," and it doesn't say how many tokens or dollars that is. Third-party models such as Claude, GPT, and Gemini aren't in the pool. On Premium and above they are charged "at the model's API list price from the first request," paid from prepaid credit you top up. The Cloud Agents page says the same thing from the other direction: Cloud Agents are charged at API pricing for the selected model.

The current docs don't put a number on a heavy month. They say instead that Premium suits new agent users, Super suits daily agent users, and Ultra suits people running several agents or automations.

The line that matters is on the team plans. From the same pricing docs: "On Teams and Enterprise plans, third-party model requests include a Cursor Token Rate of $0.25 per million tokens. This rate applies on top of model API pricing for additional usage and BYOK usage. Individual plans don't pay it."

So on Teams, every third-party token the agent uses carries a quarter of a dollar per million on top of the provider's price, and the sentence says explicitly that this includes tokens bought with your own key. The fee is small and it is published, and it is still a per-token charge on a credential you brought yourself, which a CTO comparing invoices should know about.

On your own key more generally: Cursor supports custom API keys for OpenAI, Anthropic, Google, Azure OpenAI, and AWS Bedrock on Premium and up, and its API keys page says "custom API keys only work with chat models." Nothing on that page or the Cloud Agents page says a custom key runs a Cloud Agent. Treat it as unsupported until Cursor says otherwise.

The editor. Cursor is one of the best places to write code with a model, and Cloud Agents share its rules, its configuration, and its habits, so a task you hand off behaves like a task you would have done yourself in the same repository. You can be halfway through a change in the editor and send the rest to the cloud without leaving the window, and no standalone agent can offer that.

It is also one configuration. Rules, skills, hooks, and MCP servers are set up once and apply locally and in the cloud. A team that has already invested in Cursor's conventions gets its cloud agents for the price of a plan upgrade and no new vendor in the security review.

And the pricing is mostly legible: third-party models at API list price, every per-token rate published, and a monthly cap on additional usage that you set. The one number missing is the size of the weekly pool for Cursor's own models. Compared with credit units and unpublished quotas elsewhere, it is still a relief.

A subscription you already pay for does nothing here. If your developers hold Claude Max or ChatGPT plans, Cursor's usage pool cannot draw on them; the model bill is Cursor's and lands on Cursor's invoice at API price, plus the token rate on Teams. For a team that bought subscriptions precisely so the model line would be paid once, that is buying the same tokens twice.

There is no free way to try one on a real ticket, because Free and Plus exclude Cloud Agents. Cursor is entitled to that choice, and it is friction for someone who wants to see a pull request before asking for a card.

The GitHub issue path is a comment, not a handoff. A label or an assignment lets someone start work without composing anything; a mention asks them to write a prompt in the issue. It works, and it is a different gesture.

And the review loop wants either a mention or a prompt that asked the agent to watch its pull request. A changes-requested review on an agent that wasn't told to watch does not, as far as the docs describe, send it back to work.

If your team writes code in Cursor, use them. The integration with the editor is the product, the pricing is honest, and the Linear assignment is as good as anyone's. Most people reading this should stop here.

If your backlog lives in GitHub Issues and you want a label-driven handoff, or your developers already pay for Claude or ChatGPT and you want those subscriptions to do the work, or you want the agent to act on every review without being asked to watch for one, that is the gap. Those are the three things Sinatra does differently: a Sinatra label on a GitHub issue or an assignment in Linear, your key or your subscription running every hosted task with nothing added, and a follow-up commit from a changes-requested review. Pro is $20 a month for the whole workspace, with 3 agents at a time, plus 10 waiting, and the free tier is 1 agent at a time, plus 1 waiting, with no daily limit, on your own key or subscription, no card. It has no editor, so for most teams it runs alongside Cursor rather than instead of it. The full comparison has the table and the sources, and the alternatives hub has the same table against ten other agents.

If you want to run the same tickets through both and compare the pull requests, Cloud Agents and Sinatra do not collide: one answers to @cursor and a Linear assignee, the other to a label and an assignee of its own. Start free and put your own key or subscription behind it, then read the two invoices side by side.