agent-browser

Agent browser

This example shows how to give an AI assistant in a WPF application a browser that stays within the limits the application sets. The assistant, Claude, drives a visible BrowserView while the user watches. The application enforces the limits in-process through DotNetBrowser handlers, independently of the model:

• a site allowlist with read-only and read-write sites, covering redirects, frames, and the requests a page loads;
• no downloads, pop-ups, printing, file dialogs, external apps, or permissions;
• human input suppressed while the assistant acts, with Take over and Hand back buttons;
• no typing into password fields: the assistant hands the page to the user to sign in;
• a dialog that shows the exact form data before anything is submitted;
• an audit log of navigations, blocked actions, tool calls, and token usage;
• an incognito profile, so nothing is inherited or kept.

The demo runs against a mock supplier portal served from the application's resources, so the scenario is the same on every machine.

The tutorial is available on the DotNetBrowser blog.

Prerequisites

• .NET 8 SDK
• Windows 10 or later
• A DotNetBrowser license key
• An Anthropic API key

Project structure

• GuardedBrowser is a class library with no UI and no model code:
  • BrowserPolicy.cs holds the allowed sites and schemes.
  • PolicyEnforcer.cs sets the navigation, network, download, pop-up, permission, print, and dialog handlers.
  • InputGate.cs decides whether the user or the assistant holds the keyboard and mouse.
  • SubmitGuard.cs cancels form submissions the user has not approved.
  • PageSnapshot.cs and snapshot.js describe the page as a numbered list of interactive elements.
  • BrowserActions.cs performs the assistant's actions with simulated input.
  • AuditLog.cs records what happens.
• AssistantAgent is a class library that connects Claude through the Anthropic C# SDK:
  • BrowserTools.cs defines the tools the model can call.
  • AssistantSession.cs runs the tool loop, pauses while the user holds the page, and asks for approval before a form is submitted.
  • SystemPrompt.md tells the model about the limits.
• AgentBrowser.Wpf is the application:
  • MainWindow.xaml shows the browser, the task box, the buttons, and the audit log.
  • ApprovalDialog.xaml shows what the assistant wants to submit.
  • Portal/ contains the mock supplier portal and the scheme handler that serves it at dnb://portal/.

Set the license key

Put your DotNetBrowser license key into the dotnetbrowser.license file in the root directory of this repository. The build copies it to the output directory of the application.

Set the Anthropic API key

The Anthropic SDK reads the key from the ANTHROPIC_API_KEY environment variable. Set it in the terminal you run the example from, or in the run configuration of your IDE:

$env:ANTHROPIC_API_KEY = "<your_api_key>"

Run the example

From this directory:

dotnet run --project AgentBrowser.Wpf

Or open AgentBrowser.sln in your IDE and run AgentBrowser.Wpf.

Click Run to start the task shown in the task box. When the assistant reaches the sign-in page, it hands the page over: sign in with user name demo and password demo, then click Hand back. Before the order is placed, the approval dialog shows the form data; click Submit to send it.