Power Digital

// AI News

GitHub Copilot browser use explained: which jobs to hand over

Alex Ng
AI News
Illustration with the headline: Copilot Browser Use Explained

In short: On 17 November 2026 GitHub put browser use into public preview for Copilot, so it can fill in forms, read pages and test web apps in a browser you are already signed in to. It is off by default, GitHub has not published pricing for it, and our view is that a developer should still set it up and review the results.

What happened

GitHub published three changes for the GitHub Copilot app on 17 November 2026, tied to Microsoft Ignite. First, Azure DevOps repositories are supported in public preview. Second, GitHub Enterprise Server moved from an opt-in preview to general availability. Third, and the one most relevant to non-developers, browser use arrived in public preview.

Browser use lets Copilot work inside a Chromium-based browser, including Google Chrome and Microsoft Edge. GitHub says it can fill out forms, pull information from web pages, and test or debug the web app you are building, in the browser where you are already signed in. It also works alongside computer use, which must be enabled first in the Copilot app. GitHub's separate post on the feature lists four places to use it: the Copilot app, Copilot CLI, VS Code and JetBrains IDEs.

Key details

  • Status: public preview. GitHub's documentation says the feature is subject to change.
  • Default: off. The documentation says it is disabled by default.
  • Requirement: the GitHub Copilot extension for Chrome. The Ignite changelog post says it runs in an isolated tab group, and GitHub's browser post says the extension should show a Connected state. For Edge, the documentation says Google Chrome must also be installed.
  • Turning it on in the Copilot app: Settings, then Computer Use, then switch on both Enable Computer Use and Enable Browser Use. Computer use has to be enabled first.
  • Turning it on elsewhere: in VS Code and JetBrains, Settings, then Enable Browser Use. In Copilot CLI, enter /plugin, enable browser use, and start a new session.
  • Permissions: GitHub says permission levels are managed in the extension's settings.
  • Admin control: enterprise administrators can set features.browserUse to false in managed settings so users cannot enable it. If it is true or omitted, users decide for themselves.
  • Price and plans: Pricing, plan eligibility and usage limits have not been published. GitHub has not announced Singapore availability.

Alongside this, the Azure DevOps preview lets the app clone a repository, create a branch, make changes and open a pull request. It is switched on under Settings, then Experimental. Cloud sessions are not available for Azure DevOps or for GitHub Enterprise Server repositories. The Enterprise Server release needs Copilot app version 1.2.0 or later.

What GitHub says it is for, and what it warns about

GitHub's documentation gives three example uses: testing a website and recording the results in an online spreadsheet with a drafted summary email, filling out forms, and going to a page to retrieve information. It adds that you can ask Copilot to show an email before sending it.

It is also direct about limits. Web interfaces change, so the agent can pick the wrong control, type into the wrong field, or struggle with dynamic pages and long workflows. Vague instructions or unexpected page content can cause actions that affect your data or connected accounts. Pages may also show sensitive information, including information about other people. GitHub's guidance is not to ask it to submit real credentials, not to allow changes you have not reviewed, and to review the sites, actions and results.

How it compares

The useful comparison for a business is against what you do today:

  • Against a person doing the task: the agent works in your signed-in session, so it can reach whatever you can reach. That is the benefit and the risk in one sentence.
  • Against a scripted test: a written test repeats the same steps exactly. An agent follows instructions in plain language and can adapt, but GitHub says it may struggle with dynamic pages, complex workflows and timing changes.
  • Against a custom agent: Copilot's browser use is a general tool inside a developer product. A purpose-built agent for one process can be limited to one site, one account and a short list of actions. That is a different design choice, not a ranking.

What it means for businesses

The Copilot app is aimed at people who write software, and the setup steps assume you can install a browser extension and change settings. A non-technical owner can still use it, but the safe configuration involves decisions about accounts and access. The table below is Power Digital's view, reasoned from the facts above.

Which jobs to hand over, and which still need a developer

JobOwner or non-technical staffSmall dev teamWhy (our view)
Click through your own website and list what looks brokenReasonable, with reviewYesMatches GitHub's testing example. Use a staging site or a logged-out session, and read the report before acting on it.
Check that a contact or booking form submits correctlyOnly on a test formYesA live form could send real messages or create real bookings. Use dummy data and a test inbox.
Copy product or price information from supplier pages into a sheetReasonable, with spot checksYesMatches the "retrieve information" example. GitHub warns the agent can pick the wrong control, so compare a sample against the source.
Fill in repetitive forms on a portal where you are signed inNot yetCautiouslyIt acts as you. Mistakes may affect real accounts, and GitHub says not to give it real credentials.
Anything that sends money, submits tax or regulatory filings, or changes customer recordsNoNo, not in previewKeep a person doing these.
Draft an email summarising test resultsYesYesAsk it to show the email first, as the documentation suggests, and send it yourself.
Debug a web app in real timeNoYesNeeds someone who can read the code and judge whether the fix is right.
Anything involving customer personal data on screenNoOnly after a policy decisionPages can show information about other people. Whether that is sent onward is a data protection question to settle first.

Settings and decisions to make before switching it on

  • Decide who may enable it. If you manage an enterprise, features.browserUse set to false blocks it. If you leave it out, every user decides alone.
  • Use a separate browser profile. GitHub says the agent works where you are signed in. A profile with only the test sites logged in limits what it can reach.
  • Check the extension's permission levels. GitHub has not listed the permission levels. Read what is offered on your own install, and start with the most restrictive one.
  • Use test accounts and dummy data. This follows GitHub's advice not to submit real credentials.
  • Decide what must be approved by a person. Anything that sends, pays, deletes or changes records should have a human sign off. GitHub's own advice is to review sites, actions and results before allowing changes.
  • Keep a record. Note which tasks were given to the agent, on which sites, and who reviewed the output.
  • Ask about cost. Pricing for this feature has not been published, so check your Copilot plan before relying on it.
  • Remember it is a preview. Behaviour and settings may change, so do not build a daily routine on it yet.

If your staff handle customer information on screens an agent could see, the controls above overlap with the points in our checklist for Singapore SMEs adopting agentic AI, which covers approvals, limits on what an agent can do and human oversight.

Our take

This is Power Digital's view. Browser use is a sensible addition for developers who want an assistant to click through the app they are building and report what breaks. For a business owner, the first safe uses are read only: checking your own site, collecting public information, drafting a summary. The risky uses are those where the agent acts as you on a live account.

The gaps matter. GitHub has not published pricing, has not named the permission levels, and says the agent can click the wrong thing. Until those are clear, we would keep a person on every step that changes data or touches customers, and we would not use it on pages showing personal data.

If you want an agent limited to one process, one site and a fixed list of actions, that is a different build from a general tool in a developer app. Our AI agent development service covers that kind of scoped agent, and we also build and test the sites and tools such agents work with through our web application development service.

Sources

Tags

#ai agents #browser use #GitHub Copilot #singapore business #developer tools

// what we build

Explore Our Services

// AI News

Insights & Resources

Read more →
Back to Articles