# From the browser agent to an agent that works locally

> In the browser, an agent lives in a tab and works with whatever you paste. Locally, it works inside the project folder: it reads, writes and organizes files alongside you. The folder can be shared with your team, and every change is kept as a version.

Author: Felippe Oliveira · Published 2026-11-18
URL: https://blog.felippeoliveira.com/en/agent-that-works-locally/

An agent that works locally is installed on your computer and works inside a folder: a project's, a department's, a recurring task's. In the browser, the agent stays in a tab and only sees what you paste or upload. The difference is big: AI stops talking about the work and starts doing it with you.

## What changes when the agent leaves the browser?

It starts working where the content already is. Instead of copying a passage, pasting it into the chat and copying the answer back, you ask, and the agent opens the files, makes the change and shows you what changed.

| | Browser agent | Agent that works locally |
|---|---|---|
| Where it works | Inside a tab | In the project folder |
| Your files | Only what you upload | Reads, writes and organizes them |
| Output | Text for you to copy | The work done in place |
| Team | Everyone in their own chat | Everyone in the same folder |
| History | Stays in the chat | Every change becomes a version |

## Why work in a project folder?

Because the folder holds everything the work needs, and the agent always starts from that content. Whatever it produces stays there for the next round.

```text
product-launch-campaign/
├── brief.md
├── research/
├── copy/
├── budget.xlsx
└── agent-instructions.md
```

So the work gets better with each round: you ask, review, adjust, and the result is saved. The team's instructions live in a file the agent always reads, so nobody has to repeat the same request.

<figure class="diagram">
  <img class="diagram-light" src="/diagramas/project-folder.svg#light" alt="Diagram with the project folder in the center and four parts around it: you, who ask, review and decide; the agent, which reads, writes and organizes files; the team, which works in the same folder; and versions, which keep every change." loading="lazy" decoding="async" />
  <img class="diagram-dark" src="/diagramas/project-folder.svg#dark" alt="" aria-hidden="true" loading="lazy" decoding="async" />
</figure>

## How does the team work together?

The same folder can be shared with your team or with whoever is part of the project. Each person works with their own agent on the same content.

With version control, every change is recorded: what changed, who changed it and when. A folder with that history is called a repository. You can compare versions and roll back when a change, yours or the agent's, did not turn out well.

```text
Nov 18  Agent: summarizes the interviews in research/summary.md
Nov 18  You: adjust the tone of the main copy
Nov 17  Agent: creates the first version of the brief
```

## Does the model run on the computer too?

In most cases, no. The agent works on the computer, but the model, which generates the responses, runs at the provider: the service that delivers the model over the internet. Whatever the agent reads to respond goes there. That is why it pays to choose carefully which folders it can open.

## How do you use it today?

In the browser, in the desktop app or on your phone? Do you already use your phone to direct the agent running on your computer?

## Frequently asked questions

### Do I need to know how to code to use a local agent?

No. You ask in plain language, and the agent takes care of opening files and running commands. A basic grasp of folders and versions helps.

### What kind of folder works?

Any working folder: a project, a department, a recurring task. It works best when the content is in documents, spreadsheets and text files the agent can read.

### Is a local agent riskier than a browser one?

It reaches more things, so a mistake has bigger consequences. That is why the agent has permissions: allowed folders, commands that ask for confirmation and blocked actions. Version control helps undo what did not turn out well.
