Agent Browser

Agent Browser (CLI)

Use this skill when we need to verify UI behavior ourselves (login flows, redirects, error toasts) without guessing.

Install / Setup (If agent-browser Is Missing)

1. Check if the CLI is available:

command -v agent-browser && agent-browser --help | head

2. Install Agent Browser CLI (pick one):

macOS (Homebrew):

brew install agent-browser

Linux:

• First install Node.js + npm (examples):

# Debian/Ubuntu
sudo apt update && sudo apt install -y nodejs npm

# Fedora
sudo dnf install -y nodejs npm

# Arch
sudo pacman -S --needed nodejs npm

If your distro Node.js is too old, use a version manager (nvm/fnm/asdf) and install a current Node LTS.

All platforms (npm, recommended/fallback):

npm install -g agent-browser

Windows note:

• If you don't have Node.js yet, install it first (PowerShell):

winget install OpenJS.NodeJS.LTS

3. Install the browser runtime the CLI uses (one-time per machine/user):

agent-browser install

Linux-only note:

• If the browser fails to launch due to missing system libraries, try:

agent-browser install --with-deps

4. Sanity check:

agent-browser open https://example.com/
agent-browser snapshot -i
agent-browser close

Artifacts (Cross-Platform)

Prefer saving evidence inside the repo so it can be attached/shared predictably:

• Default: .agent/artifacts/agent-browser/<run-id>/
• Include: screenshots (.png) and snapshots (.txt)

Example (macOS/Linux):

run_id="oidc-$(date +%Y%m%d-%H%M%S)"
out_dir=".agent/artifacts/agent-browser/$run_id"
mkdir -p "$out_dir"
echo "$out_dir"

Example (Windows PowerShell):

$runId = "oidc-" + (Get-Date -Format "yyyyMMdd-HHmmss")
$outDir = ".agent/artifacts/agent-browser/$runId"
New-Item -ItemType Directory -Force -Path $outDir | Out-Null
$outDir

Quick Start (OIDC Debug)

The core loop is: open -> snapshot -i -> interact (prefer refs like @e2) -> re-snapshot -i -> capture evidence.

macOS/Linux (bash)

1. Start a clean session, open the app, and capture baseline evidence:

run_id="oidc-$(date +%Y%m%d-%H%M%S)"
out_dir=".agent/artifacts/agent-browser/$run_id"
mkdir -p "$out_dir"

agent-browser close || true
agent-browser --session oidc open https://example.com/   # replace
agent-browser --session oidc wait 1000
agent-browser --session oidc screenshot "$out_dir/01-home.png"
agent-browser --session oidc snapshot -i > "$out_dir/01-home.snapshot.txt"

2. Prefer clicking by ref from the snapshot (fallback to semantic find if needed):

# Look in 01-home.snapshot.txt for: button "Sign in" [ref=e2]
agent-browser --session oidc click @e2

# Fallback if refs aren't obvious:
# agent-browser --session oidc find role button click --name "Sign in"

agent-browser --session oidc wait 3000
agent-browser --session oidc get url
agent-browser --session oidc screenshot "$out_dir/02-after-login-click.png"
agent-browser --session oidc snapshot -i > "$out_dir/02-after-login-click.snapshot.txt"

3. If the flow fails, capture high-signal diagnostics:

agent-browser --session oidc errors
agent-browser --session oidc console
agent-browser --session oidc network requests

Windows (PowerShell)

$runId = "oidc-" + (Get-Date -Format "yyyyMMdd-HHmmss")
$outDir = ".agent/artifacts/agent-browser/$runId"
New-Item -ItemType Directory -Force -Path $outDir | Out-Null

agent-browser close
agent-browser --session oidc open https://example.com/   # replace
agent-browser --session oidc wait 1000
agent-browser --session oidc screenshot "$outDir/01-home.png"
agent-browser --session oidc snapshot -i | Out-File -Encoding utf8 "$outDir/01-home.snapshot.txt"

# Click by ref from the snapshot, e.g. @e2
agent-browser --session oidc click @e2
agent-browser --session oidc wait 3000
agent-browser --session oidc get url
agent-browser --session oidc screenshot "$outDir/02-after-login-click.png"
agent-browser --session oidc snapshot -i | Out-File -Encoding utf8 "$outDir/02-after-login-click.snapshot.txt"

agent-browser --session oidc errors
agent-browser --session oidc console
agent-browser --session oidc network requests

Interaction Guidance (Stability First)

• Prefer refs from snapshot -i (for example @e2) for deterministic selection.
• If you can't use refs, prefer find role ... over CSS selectors; it is more stable across UI changes.
• Use --session <name> consistently so multiple investigations can run without stepping on each other.
• Re-run snapshot -i after navigation or DOM changes (refs are tied to the current page state).
• Use wait <ms> only as a last resort; prefer waiting on UI cues (for example, find role heading or find text) when possible.
• Use agent-browser close between runs if you need to reset global state.

Security / Privacy Notes

• Avoid capturing credentials, secrets, or customer data in screenshots/snapshots.
• When you must test auth flows, prefer test accounts and scrub/redact artifacts before sharing outside the immediate team.