What the 6 October note says
Cursor’s 6 October release note describes viewing and replying to local agents in its iOS app. It says the computer must remain awake and online, pairing is approved on desktop, and Remote Control is enabled by default outside Enterprise. It also says cloud agents are not required.
The current docs draw a different boundary
The current iOS documentation, read 10 October, says the agent loop runs in Cursor’s cloud while terminal commands, file changes, tests and git run on your computer. Conversation state, tool results and model context cross that boundary. It specifies desktop 3.9.8 or later, Cloud Agents access and cloud data storage, with team administrators enabling access.
Those descriptions do not fully agree about cloud requirements or default controls. We could not establish whether they describe different implementations, account rollouts or documentation changes. Treat the installed client and your organization’s effective settings as unresolved checks; the release headline alone cannot establish that all processing stays on the laptop.
A small handoff check you can record
For someone supervising a development task away from the desk, the useful evidence is a continuous session with a clear execution location. The following is a proposed check, not a Fewertools test result.
- Write down the starting state. Record desktop version, account plan, workspace type and the exact access controls shown. Keep screenshots free of credentials and private repository contents.
- Use a disposable project. Give the agent a harmless task such as changing a local text fixture and reading its contents back. Record the starting file and expected result before handing off.
- Observe one round trip. Send a follow-up from the phone, then compare the transcript, file diff and terminal output on desktop. Note whether the same session continues and which approval prompts appear.
- Record the failure path. On the disposable task, note what the phone reports if the computer becomes unavailable. A message that was submitted is a different observation from a tool call that finished.
A useful receipt would contain the versions, task, timestamps and resulting diff. It would still not prove the handling of every data type or every enterprise policy. No client pairing, privacy-setting change or native agent run was performed for this report.
Read this alongside
The Cursor profile covers the wider product. Our Copilot local-model update examines a related distinction: selecting a local runtime and choosing a network policy are separate decisions.
Sources and scope
Official sources read 10 October 2026. Facts above are attributed to the provider; the evaluation worksheets are proposed checks. This article does not report a native product test or assign a tool score.