Manus vs APX: where does agent work happen?

“Does it run locally?” is a useful question when choosing an AI agent, but it is not specific enough. The answer can differ for file operations, the agent’s planning, model inference and saved history. Ask where each part happens, what leaves the device and what continues when your computer is off.

This article is written by APX using the linked documentation checked on November 8, 2026. It is not a hands-on performance or security comparison. Product behavior and terms can change.

Manus: a choice of work locations

Manus Studio’s documentation describes a default cloud sandbox, a persistent Cloud Computer option and a Work locally mode on a user’s computer. It says local work can use selected files, terminal tools and hardware while Studio is running, but also says Manus plans in the cloud. Some features upload files, and the documentation says a task can temporarily switch to the cloud when a feature does not work locally. Review the current disclosure and permissions for the task you intend to run.

The same guide describes permission modes and notes that its two approval modes limit changes, not all reads accessible to the user account. Local execution should not be treated as proof that nothing leaves your computer.

APX: project work on your computer or server

The APX agents guide describes project-specific roles and handoffs, with APX running on your computer or a server you control. Its models guide lists cloud providers and local model options. A cloud provider receives the prompts needed for a request; hosting APX yourself is not the same as using an offline model.

APX’s autonomy guide describes its permission modes and their boundaries. Read the mode and tool permissions you actually select; neither a product name nor a local folder is a substitute for reviewing an approval.

A fair test for your own work

Try the same small report in both with non-sensitive sample files. Record:

1. Which folder the agent reads and which files it changes.
2. Which operations ask for approval and how you inspect them.
3. Where planning and model requests run, and which provider terms apply.
4. What happens when the computer sleeps or the network is unavailable.
5. The actual account and model usage shown for the test, including any separate infrastructure cost.

Manus may fit when its integrated deliverable workflow and choice of cloud or local workspace suit you. APX may fit when you want project-specific agents and a task board on your own installation with a choice of providers. Neither claim establishes which will produce the better report; inspect the output and its sources yourself.