AgentComparison
Anysphere (Cursor) · product · coding agent (editor and cloud agents)

Cursor cloud agents

An AI code editor whose cloud agents work on branches and open pull requests.

Group: Coding agents

Cursor is an AI code editor that also runs cloud agents: isolated virtual machines that clone your repository, work on a separate branch, test the change and push it for review. Agents can be started from the editor, web, mobile, Slack, GitHub, Linear or the API.

Documentation researched, not testedChecked 22 October 2026Vendor documentation, not a test

The decision in plain English

Use cloud agents for bounded repository tasks that you can review as draft pull requests. They auto-run terminal commands, unlike local approval modes, and internet access is on by default. Before starting, restrict repositories and outbound domains, choose secret types deliberately, verify Privacy Mode, and check usage budgets. Privacy Mode does not mean the run leaves no stored transcript or repository snapshot.

Best fit

Developers who want background work on repository branches and can review diffs, tests and artifacts before a human merges the draft pull request. Configure restricted egress and minimum repository access first.

Poor fit

Sensitive repositories without reviewed cloud-storage and network policies, workflows needing approval for every command, or users who assume a fixed subscription eliminates on-demand charges. Self-hosted tools still use Cursor's cloud agent loop.

Documented capabilities

Vendor statements, not verified task outcomes.

  • Managed VM repository work with draft PR and artifact handoff Source · checked 2026-10-22
  • Managed-cloud egress modes with user, environment and team policy Source · checked 2026-10-22
  • Environment Variable, Runtime Secret and Build Secret scoping Source · checked 2026-10-22
  • Managed or self-hosted tool execution with the agent loop in Cursor cloud Source · checked 2026-10-22

Permissions and human control

  • An admin installs the Git provider app and selects repositories; each user connects their own Git identity. Cursor documents that the agent cannot reach repositories the triggering user could not already reach.
  • Cloud agents run terminal commands without per-action approval and open draft pull requests for human review before merging. Do not apply the local Auto-review, Allowlist or Run Everything settings as cloud-agent gates.
  • Internet access is on by default. Managed cloud users, saved environments and teams can choose Allow all, Default + allowlist or Allowlist only; admins can lock policy. Scope exact destinations, not broad storage-domain wildcards.
  • Environment Variable secrets are visible to the agent. Runtime Secrets are redacted from tool results, transcripts and commits, but still exist as environment variables and are visible to people using the environment Terminal. Build Secrets are available only during Docker builds.
  • Secret redaction is not credential revocation. Remove stored secrets/tokens when finished, and check transcript and snapshot retention separately. Network controls for managed agents do not govern self-hosted workers.

Documented controls do not prove that an agent always follows them.

Limits and open questions

  • Paid cloud-agent access is distinct from Hobby's limited local Agent requests; removed the Free cloud-agent filter tag
  • No per-command approval: cloud agents auto-run terminal work, with prompt-injection and exfiltration risks described by Cursor
  • Default internet access can be restricted; managed-cloud egress settings do not extend to self-hosted machines
  • Conversation history is indefinite by default; deleting a transcript and artifacts does not immediately delete environment snapshots
  • Snapshots expire after 30 days of inactivity; starting or resuming extends that window
  • Cursor privacy documentation differs on whether all cloud runs enforce Privacy Mode; verify the actual starting state before sensitive use
  • Vendor documentation only; no hands-on reliability test or security audit

Privacy and security

With Privacy Mode on, Cursor says cloud-agent code, prompts and responses are not used for training. This is not zero storage: conversation history is indefinite by default and can be deleted with its artifacts; environment snapshots cannot be deleted on demand and expire after 30 days of inactivity, extended on start/resume; Enterprise teams can cap conversation retention with retention policies. Secrets and tokens remain until removed. The security overview says cloud agents run in Privacy Mode, while Secrets & Network still describes disabled-mode collection and warns that enabling it mid-run does not change that run. Treat this documentation mismatch as a reason to check the actual starting policy. Most models have zero-data-retention agreements, but the overview notes exceptions; this check did not verify each model. These are vendor statements, not an independent audit.

Feature-by-feature: what is documented and what is not

Official sources checked 22 October 2026. A documented feature is a vendor statement, not proof of success.

FeatureEvidence statusDetail and limit
Repository accessAdmin and user scopeGit provider app repository selection and triggering-user access both apply. Protected Git Scopes and a blocklist can narrow access. Source
Command approvalAuto-run in cloudCloud agents run terminal commands without per-command approval. Local Run Modes are not cloud gates; review the draft pull request before merging. Source
Network policyConfigurable, not absoluteAllow all, Default + allowlist and Allowlist only apply to managed cloud. Essential Cursor/SCM domains remain accessible even in allowlist-only mode. Source
Secret typesDifferent exposureEnvironment Variables are agent-visible. Runtime Secrets are redacted from outputs/commits but visible in Terminal; Build Secrets are build-only. Source
Transcript deletionSeparate from snapshotsConversation history is indefinite by default and deletable. VM snapshots cannot be deleted on demand and expire after 30 days of inactivity; each start or resume extends the window. Snapshot rule rechecked 12 November 2026. Source
Privacy ModeDocumentation mismatchSecurity overview says cloud runs use Privacy Mode, while the configuration reference still describes collection when it is disabled and says enabling it mid-run does not change that run. Check the starting policy and model-specific exceptions. Source
Self-hosted workNot a local agent loopMy Machines and Enterprise Team Pools move tool execution to owner hardware, but the agent loop stays in Cursor cloud. Managed-cloud egress controls do not cover those workers. Source
Paid accessUsage limits applyCloud agents are listed on paid Individual and Teams plans, not Hobby. Included model usage has limits; on-demand usage is billed in arrears and tax is extra. Source

How to start with Cursor cloud agents narrowly

  1. Use one non-sensitive repository and a task with acceptance tests.
  2. Restrict Git app scope, select the runtime and configure exact outbound domains.
  3. Check starting Privacy Mode, secret type, transcript/snapshot retention and usage budget.
  4. Review the draft pull request, tests and artifacts before a human merges it.

Official product/setup source. These are documentation-based steps, not a hands-on setup test.

Common questions about Cursor cloud agents

What is Cursor cloud agents for?

Developers who want background work on repository branches and can review diffs, tests and artifacts before a human merges the draft pull request. Configure restricted egress and minimum repository access first. The vendor describes it as: Cursor is an AI code editor that also runs cloud agents: isolated virtual machines that clone your repository, work on a separate branch, test the change and push it for review. Agents can be started from the editor, web, mobile, Slack, GitHub, Linear or the API.

Who should look elsewhere?

Sensitive repositories without reviewed cloud-storage and network policies, workflows needing approval for every command, or users who assume a fixed subscription eliminates on-demand charges. Self-hosted tools still use Cursor's cloud agent loop.

How much does Cursor cloud agents cost?

Cursor's pricing page (read 22 October 2026) lists Free at $0, Plus at $8/month, Premium at $30/month (marked most popular), Super at $100/month, Ultra at $200/month, Teams from $30/user/month and custom Enterprise. This replaces the earlier Individual $20/month and Teams $40/user/month structure. Plan inclusions are usage-based; third-party models are billed separately by usage. Re-read 28 October 2026: feature table lists 'Cloud agents & automations' on Premium ($30/month) and above - Free and Plus ($8) are excluded. Page now also lists code hosting via Origin, AI code reviews via Bugbot, billing via X Subscription Corp and a usage pool shared with Grok; third-party models remain at API list price with spending caps.

What can it access and what needs my approval?

An admin installs the Git provider app and selects repositories; each user connects their own Git identity. Cursor documents that the agent cannot reach repositories the triggering user could not already reach. Cloud agents run terminal commands without per-action approval and open draft pull requests for human review before merging. Do not apply the local Auto-review, Allowlist or Run Everything settings as cloud-agent gates. Internet access is on by default. Managed cloud users, saved environments and teams can choose Allow all, Default + allowlist or Allowlist only; admins can lock policy. Scope exact destinations, not broad storage-domain wildcards. Environment Variable secrets are visible to the agent. Runtime Secrets are redacted from tool results, transcripts and commits, but still exist as environment variables and are visible to people using the environment Terminal. Build Secrets are available only during Docker builds. Secret redaction is not credential revocation. Remove stored secrets/tokens when finished, and check transcript and snapshot retention separately. Network controls for managed agents do not govern self-hosted workers.

What happens to my data?

With Privacy Mode on, Cursor says cloud-agent code, prompts and responses are not used for training. This is not zero storage: conversation history is indefinite by default and can be deleted with its artifacts; environment snapshots cannot be deleted on demand and expire after 30 days of inactivity, extended on start/resume; Enterprise teams can cap conversation retention with retention policies. Secrets and tokens remain until removed. The security overview says cloud agents run in Privacy Mode, while Secrets & Network still describes disabled-mode collection and warns that enabling it mid-run does not change that run. Treat this documentation mismatch as a reason to check the actual starting policy. Most models have zero-data-retention agreements, but the overview notes exceptions; this check did not verify each model. These are vendor statements, not an independent audit.

Has AgentComparison tested it?

No. This record is built from the vendor's own documentation, checked 2026-10-22. Vendor claims are attributed, not confirmed.

How to evaluate Cursor cloud agents yourself

  1. Pick one low-risk task that matches its documented capabilities: managed vm repository work with draft pr and artifact handoff.
  2. Write the expected output and what counts as failure before you start.
  3. Connect the minimum access needed, and check approval settings first.
  4. Run it twice and compare, then review every action it took.
  5. Check cost after any free allowance. Unknown price is not free.

Related

Documented use cases: Software development and repository work.

Sources and history

Change history

Documentation overview expanded with current pricing/access sources. No hands-on results claimed.

Report a correction