WebBrowser puts the application in front of you and your agent
WebBrowser is a live Chromium surface inside the Canvas. A connected CodingAgent can observe, interact, wait, debug, and test the page you are also seeing.
See the WebBrowser demonstration in Kavor (opens in a new tab) in another tab while you read this guide.
The browser as a development tool
To develop a web application, a CodingAgent needs more than files. It needs to open the application, interact with it, wait for real states, and investigate what happened in the browser.
With a CodingAgent + WebBrowser Connection, the agent can use a broad operation surface mediated by Kavor:
- observe: read page state, obtain an accessibility snapshot, and capture screenshots;
- interact: click, fill fields, insert text, press keys, select options, check controls, scroll, drag, upload files, and respond to dialogs;
- synchronize: wait for a selector, text, URL, load state, network idle, or a specific condition;
- debug: inspect console messages, network requests, and retained response bodies;
- test: reproduce a bug, run an E2E flow, and preserve visual evidence of the result;
- isolate scenarios: block, continue, or fulfill network responses during a controlled test;
- organize pages: open, select, and close tabs, follow downloads, and handle authentication challenges.
The goal is not to hide the browser behind automation. It is to make state and actions verifiable while the agent works.
A practical loop for a web application
Start with a WebBrowser and a CodingAgent in the same component. If the application runs locally, connect the Terminal that starts the server as well. A small loop can be:
File / Specification — CodingAgent — WebBrowser
│
Terminal
Ask the agent to:
- observe the page before acting;
- reproduce the failing path;
- collect console, network, or screenshot evidence;
- change the code within the defined scope;
- wait for the new state and repeat the flow;
- record the result and remaining risks in a Sticky Note.
A useful starting prompt is:
Open the connected application in WebBrowser. First observe the page and reproduce the flow without changing code. Then describe the likely cause, propose the smallest change, and validate the complete path with visual and console evidence.
The same browser for the human
You can also use WebBrowser as an ordinary page inside the Workspace: open documentation, watch a video, or leave a reference page open while agents work.
YouTube and other common pages are natural uses. Streaming services such as Netflix may require authentication, DRM, permissions, or system-specific conditions, so Kavor does not promise playback for any particular service.
A shared surface, not an invisible browser
The agent and the human share the same live page. This has useful consequences:
- you can see the actions and intervene;
- the agent has no private window hiding what it is doing;
- persistent tabs and presented pages belong to Kavor's own browser profile, shared among its WebBrowsers;
- extensions, history, and cookies from your external Chrome are not automatically reused;
- page content is untrusted and cannot redefine the agent's instructions.
WebBrowser is not a remote browser service or a general permission to operate the machine's filesystem.
Limits and care
Observe first, then act
Before controlling a page, the agent should read its current state. Snapshot references are temporary and can expire after navigation or DOM changes. When that happens, take a new snapshot instead of insisting on the old reference.
Sensitive actions remain human-owned
CAPTCHAs, passkeys, site permissions, certificates, and authentication prompts may require your intervention. The agent can detect or wait for these situations, but must not pretend that a human action happened.
Network rules are temporary
Blocks and fulfilled responses apply to the controlled scenario and page generation. Clear the rules when the test ends; they are not a permanent application configuration.
The development profile accepts certificates
To reach local servers and development environments, the dedicated profile accepts self-signed, expired, and private authority certificates. This also reduces protection against a hostile network presenting an invalid certificate. Use the profile carefully before authenticating with sensitive services.
A page can remain live off-screen
When you switch Workspaces in the same window, Kavor preserves the presented page. A connected CodingAgent can continue operating that state even while another Workspace is visible. Closing the tab, deleting the Node, or ending the session ends that continuity.
Frames and pages may be partial
An accessibility snapshot may omit cross-origin frame content. A network response may have been discarded once the bounded history advanced. The agent should report observed evidence, not invent what it could not inspect.
Connection and Guardrail
The direct CodingAgent + WebBrowser pair places the browser in the agent's reachable component. The graph can include a Specification, Files, Terminal, Sticky Note, and other CodingAgents through valid paths, but visual proximity or a message mention does not create access.
WebBrowser does not have its own specific Guardrail today. This does not remove limits from resources in the same graph: a read-only File remains read-only, a Terminal session keeps its controls, and a Specification still follows its lifecycle.
Continue
- Consult the Connections matrix for the CodingAgent + WebBrowser contract.
- Read how CodingAgents see and build the Canvas.
- Combine browser, code, and evidence in your first loop.
- Use a Specification to define expected behavior before testing the application.