Problem
The agent-browser skill (v0.27.0) bundled with Command Code doesn't work reliably on Windows, making it effectively impossible to browse websites from within a session.
What we observed
1. URL mangling
Passing a full URL like https://www.google.com causes agent-browser to navigate to https://https/://www.google.com. Only bare domains like google.com resolve correctly — but even then the page load times out (30s+), even though the page content does eventually load.
2. Snapshot refs don't match docs
The skill guide and --help output both show @eN syntax for refs (e.g. agent-browser click @e4). In practice:
agent-browser click @e4 → always returns "Element not found"agent-browser click e4 (without @) → returns "✓ Done" but the click has no visible effect on the page
We're not sure if this is a parsing bug, a stale ref issue, or something else.
3. Interactive elements can't be clicked
On Google's cookie consent dialog, the buttons ("Alle ablehnen", "Alle akzeptieren") appear correctly in the snapshot with refs. However, none of these approaches actually click them:
@eN ref (element not found)eN ref (says done, nothing happens)find role button click --name "..." (element not found)- CSS selectors like
button:has-text('...') (element not found)
eval with JS .click() (returns null, no effect)
The page remains unchanged after every attempt.
4. agent-browser doctor opens a visible Chrome window
Running agent-browser doctor unexpectedly opens a headed Chrome window on the desktop. This is surprising for a diagnostic command, especially since the default mode is supposed to be headless.
5. agent-browser install was required
Chrome wasn't installed for agent-browser initially (only system Chrome existed). Running agent-browser install downloaded Chrome 149.0.7827.54 to ~\.agent-browser\browsers\. This may be expected, but the error messages before running install weren't clear about what was missing.
Environment
- Windows 10/11, win32-x64
- Node.js v22.17.0
- agent-browser 0.27.0
- System Chrome at
C:\Program Files\Google\Chrome\Application\chrome.exe
- No proxy, no custom agent-browser config
Why this matters
Without a working browser skill, there's no reliable way to retrieve current web information during a Command Code session. Curl and PowerShell can fetch raw HTML but can't handle JavaScript-heavy sites, consent dialogs, or structured content extraction. This is a significant gap for tasks that need up-to-date external information (pricing pages, documentation, search results, etc.).
Problem
The
agent-browserskill (v0.27.0) bundled with Command Code doesn't work reliably on Windows, making it effectively impossible to browse websites from within a session.What we observed
1. URL mangling
Passing a full URL like
https://www.google.comcauses agent-browser to navigate tohttps://https/://www.google.com. Only bare domains likegoogle.comresolve correctly — but even then the page load times out (30s+), even though the page content does eventually load.2. Snapshot refs don't match docs
The skill guide and
--helpoutput both show@eNsyntax for refs (e.g.agent-browser click @e4). In practice:agent-browser click @e4→ always returns"Element not found"agent-browser click e4(without@) → returns"✓ Done"but the click has no visible effect on the pageWe're not sure if this is a parsing bug, a stale ref issue, or something else.
3. Interactive elements can't be clicked
On Google's cookie consent dialog, the buttons ("Alle ablehnen", "Alle akzeptieren") appear correctly in the snapshot with refs. However, none of these approaches actually click them:
@eNref (element not found)eNref (says done, nothing happens)find role button click --name "..."(element not found)button:has-text('...')(element not found)evalwith JS.click()(returns null, no effect)The page remains unchanged after every attempt.
4.
agent-browser doctoropens a visible Chrome windowRunning
agent-browser doctorunexpectedly opens a headed Chrome window on the desktop. This is surprising for a diagnostic command, especially since the default mode is supposed to be headless.5.
agent-browser installwas requiredChrome wasn't installed for agent-browser initially (only system Chrome existed). Running
agent-browser installdownloaded Chrome 149.0.7827.54 to~\.agent-browser\browsers\. This may be expected, but the error messages before running install weren't clear about what was missing.Environment
C:\Program Files\Google\Chrome\Application\chrome.exeWhy this matters
Without a working browser skill, there's no reliable way to retrieve current web information during a Command Code session. Curl and PowerShell can fetch raw HTML but can't handle JavaScript-heavy sites, consent dialogs, or structured content extraction. This is a significant gap for tasks that need up-to-date external information (pricing pages, documentation, search results, etc.).