When Is Self-Hosting Cursor Cloud Agents Worth It?
Use Cursor-managed VMs for portable x86-64 jobs. Pilot My Machines or Team Pools when tool calls need your hardware, then compare total cost at the same throughput of review-accepted changes.
Make worker readiness part of the comparison
Cursor-managed Cloud Agents provide an isolated x86-64 VM, saved environments, and artifact capture. My Machines and Team Pools shift tool execution to operator-controlled hardware while the agent loop remains in Cursor's cloud. Compare those paths with a representative job that needs a clean checkout, dependency setup, and a reviewable diff. Submit it twice per path, once with the environment ready and once when the worker must be brought up from a cold state. Record queue time, setup time, and minutes until an independent reviewer can open the artifact. My Machines fits a pilot where an individual owns the host. Team Pools fits centralized fleet ownership. An illustrative worker that adds 90 seconds per cold start across 40 jobs adds an hour of aggregate delay, even if its compilation time is identical.
Choose where Cloud Agents run | Cursor DocsCharge replacement work to the deployment that needs it
Cursor's partner templates keep the Team Pool worker protocol and leave production validation and fleet operations with the operator. Compare infrastructure partners using the same worker image and one staging replacement drill. Let a disposable worker finish a job, take it out of rotation through your normal staging controls, then submit the same fixture again. Observe when a replacement becomes ready, whether the queued job completes, and whether the reviewer receives one usable result. Record billed worker time and hands-on operator minutes, including investigation of a lost or delayed result. Allocate recurring support and entitlements to monthly fixed cost, and put attempt-linked usage and worker charges in variable cost. Include failed attempts in that variable spend. A low hourly worker price can still yield a higher cost per accepted change when setup delays or failed attempts consume paid time.
Integrations | Cursor DocsProve the task travels
Cursor-hosted Cloud Agents run in isolated x86-64 cloud VMs, with saved environments and artifact capture. Before handing off a task, record the starting commit, dependencies, expected files, and the repository's existing verification command. Ask for a reviewable diff and the test output. For a task that still needs interactive diagnosis, first work through it in Cursor's editor until you can describe a checkable result.
In an illustrative trial, give the cloud run a parser bug, one input file containing a blank line, and a test that should return two records. Pass the run only if the diff fixes the count and the existing suite passes in that environment. If the suite never starts, investigate setup before judging patch quality.
Choose where Cloud Agents run | Cursor DocsRoute hardware-dependent work
Cursor Self-Hosted Machines moves tool calls to hardware you control while Cursor still runs the agent loop in its cloud. Use My Machines to test a personal worker that can stay online. Use Team Pools when an Enterprise team needs centrally owned workers or routing to a specific fleet. Confirm that your data policy permits the cloud agent loop before registering a worker.
For an illustrative ARM-only build, target an ARM worker and ask the agent to report the architecture and run the repository's existing smoke test. Pass the trial only when both outputs match expectations. If the machine reports x86-64, check worker selection. If work stalls after a personal machine goes offline, restore the worker and check its availability before retrying.
Choose where Cloud Agents run | Cursor DocsPrice a reviewed result
Cursor's Team Pool integration requires an Enterprise plan with Self-Hosted Machines enabled by an admin. For Cursor-hosted runs, verify the eligible plan, included usage, model rates and units, and overage controls in Cursor's pricing and billing views. For hosted workers, also verify the instance size, billable time, and idle policy with the host. Use those account-specific figures to estimate the incremental cost of each route.
For illustration, suppose a pilot incurs $18 in agent usage and $12 in incremental worker charges, with no change to the subscription. If 8 of 10 attempts lead to accepted changes, the cost is $30 divided by 8, or $3.75 per accepted change. Compare that with the managed pilot and track review time separately so cheap retries do not hide extra operator work.
Integrations | Cursor Docs