Guides

Local vs cloud AI agents: where do your tasks actually run?

An agent can act on your Mac while its model runs remotely. Compare local, cloud and hybrid setups by app access, inference, stored data and task requirements.

Back to Blog

A local AI agent can control apps on your computer while using a model hosted elsewhere. A cloud AI agent can operate tools in a hosted browser or virtual machine. Many setups combine local execution with remote inference, so the location of the chat window tells you little about where the work happens or where its data goes.

For computer-use tasks, follow the request through the system: which process coordinates it, which machine runs the tools, where the model makes decisions and where the resulting state is stored.

That distinction matters when the work is already open in a desktop app. A browser in the cloud needs a way to reach that project. A local tool may reach it directly, then send relevant context to a remote model.

An agent can span several machines

An agent typically repeats a loop: inspect the available information, choose an action, run a tool and inspect the result. The model and the tool do not have to run together.

Anthropic’s computer-use documentation illustrates this separation. The API returns requested actions; the application executes them in its environment and sends results back. The environment executing a click is distinct from the service producing the next decision.

Keep four locations in view:

Part of the systemWhat happens thereWhat to establish
Agent runnerCoordinates model requests and tool callsWhich process keeps the task moving, and what stops it?
Tool environmentOpens files, runs commands or controls apps and browsersWhich machine holds the app, permissions and working files?
Model inferenceProcesses context and produces the next response or actionWhich model endpoint receives the request?
Stored stateHolds conversation history, app knowledge, logs and session dataWhich records persist locally, remotely or in both places?

A fifth destination may be the service the task itself uses. A local agent uploading a file to a shared drive is deliberately transferring data. The destination belongs in the workflow description even if every decision was made on the device.

Three common arrangements have different tradeoffs

“Local” can describe the model, the tools or the entire stack. Ask which one a product means.

ArrangementWhere tools actWhere the model runsWhat it means for your task
Local tools with a remote modelYour computerA model provider’s infrastructureCan use configured local apps and files; observations and task context may leave the device
Hosted tools with a remote modelA remote browser, container or virtual machineA remote model serviceRequires the task’s apps, data and access to be available in that environment
Local tools with a local modelYour computerYour computerCan support local processing; offline operation still depends on every required tool and service

A remotely hosted environment could also call a model you host yourself. Its app operations would still happen on the remote machine.

For a concrete cloud-browser example, Browserbase documents hosted browser sessions that agents can control programmatically. That browser is the place where the web interaction happens. It needs the appropriate session and access for the job.

For local inference, Ollama’s FAQ distinguishes locally run models from cloud-hosted models and documents a local-only mode. A locally installed model client can offer both choices, which makes the selected model and configuration worth checking.

Start with the app that holds the work

Consider a small request:

Read this open TextEdit test document and summarize it without editing or saving.

With local desktop tools, the agent can inspect the document in the Mac session where it is already open, subject to the app’s capabilities and permissions. A remote model may receive a text observation and produce the summary. The app interaction happened locally; the inference happened remotely.

A hosted browser starts from a different environment. To work with that document, it needs an accessible copy, a suitable web version or an explicitly configured connection to the Mac. The document’s presence on your desktop alone does not provide that access.

The same issue becomes more consequential with a project in a professional app. Before moving the task to another machine, establish whether that machine has the required app version, files, plugins and licences. A readable export may omit the editable state the task needs.

Conversely, if the whole job is available through a web service, a hosted browser may fit without involving your desktop. Choose the environment that can reach the required state, then check the actual operations it supports.

For the distinction between API, browser and desktop access, see What is a computer-use agent?.

Where Braus runs the work

Braus gives agents tools to work in apps on your Mac. Its desktop Driver uses a virtual display within your current Mac session. The target windows and working files remain on that machine. The virtual display does not create another user account or a separate filesystem. See the background-work guide.

The desktop window-to-text pipeline also runs on your Mac. Braus turns the current interface into structured text that an agent can use. Its Memory stores interactions, outcomes and learned procedures in a local SQLite archive. Recall is still in development, and remembered steps need to be checked against the current app state.

The model connection is a separate part of this arrangement. A remote model can receive window text, task context and relevant app knowledge. Other tools, including browser tools, can return screenshots. The content sent depends on the tools and model connection used. Braus’s permissions and data documentation describes those boundaries.

A local knowledge archive tells you where that archive lives. It does not establish that every observation stays on the device. Content excluded from saved records may still appear in a live tool answer sent to the model.

Braus also supports a local MCP connection for an external assistant. Braus and the client run on the same Mac; the bridge exposes the app tools through a local connection. The client’s model connection still determines where inference happens. MCP access alone provides no guarantee of offline operation.

Availability depends on what must stay running

For local execution, the required Mac, app and agent processes must remain available. Braus’s desktop route requires an awake physical display. Its local MCP tools require Braus to stay open.

A hosted task can continue independently of your laptop only when its runner, environment and connections remain available remotely. A remote browser controlled by a script on your laptop still depends on that script. Check the whole chain before assuming that closing the lid leaves an autonomous task running.

Interruption handling matters in either setup. Stopping a task may leave an edit already applied or a submission already sent. Reconnect by observing the current state before deciding what to do next. Our article on background computer use under interruption separates these recovery questions from successful tool-call counts.

Choose the environment from the task’s requirements

For a task in an installed Mac app, establish local app access first. Check the required controls, the project’s dependencies and what information the selected model will receive. Braus is built for this setting: agents working in the apps on the Mac you already use.

For a web-only task, compare a configured local browser with a hosted browser. Login state, access restrictions, downloaded files and the destination of the result can decide which setup is practical.

For work that must keep model processing on the device, verify the inference endpoint and the complete tool path. A local model still needs enough memory and sufficient capability for the task. Tools that call remote services introduce their own data transfers. The architecture alone establishes neither task success nor performance.

Compare cost on a completed job. Include model usage, runtime or session charges, setup and recovery work where they apply. A local arrangement consumes your hardware and may still incur remote-model costs. A cloud arrangement adds its own service dependencies. A lower price per call says little about the total effort needed to obtain the correct output.

Make the first task easy to trace

Use a disposable document with distinctive test content. Name the app, the result you want and the stopping point. Start with a read-only request, then inspect the answer and the tool activity.

Before relying on the setup, be able to answer four questions: which machine performed the app operation, which model received its context, where the resulting records were stored and what would happen if the connection stopped.

For Braus, follow the setup and data guide and try the TextEdit read-only task above. Inspect the result in your own Mac session before expanding the work.

All posts