The first run explores and saves a Playwright script. Every run after replays it. No AI, no tokens. When the site changes, the AI steps back in, repairs the script (3 attempts, then it reports), and the next run is cheap again.
1st ask
minutes · AI explores
1× AI
2nd
0
3rd
0
400th
0
Ask it things like
export this month's invoices
pull new CRM leads to CSV
fill this form for every row
download last week's report
check 3 portals for changes
update 40 ticket statuses
IN CLAUDE
claude
❯/mcp__my_instance__browser_cookbook__new_recipe task: export this month's invoices
●new_recipe("export this month's invoices" )
analyzing task ... no site notes (first contact)
opening billing.acme.com/invoices
logged in via saved session
found invoice table, 12 rows
downloading INV-2041.pdf ... ✓
downloading INV-2042.pdf ... ✓
downloading INV-2043.pdf ...
browser-cookbook● working
IN CHROME
Invoices - Billing×
Dashboard
+
←→↻
🔒billing.acme.com/invoices
B
Invoices
Filter
Export All
Invoice
Date
Amount
Status
Action
INV-2041
Jul 01
$2,340.00
Paid
✓ Downloaded
INV-2042
Jul 03
$1,890.00
Paid
✓ Downloaded
INV-2043
Jul 07
$3,120.00
Paid
⬇ Downloading...
INV-2044
Jul 12
$945.00
Paid
Download PDF
INV-2045
Jul 18
$2,670.00
Pending
Download PDF
Three Actions
browse
Do it now
✓
EXPLORING
–
not saved
new_recipe
Do it + save
✓
EXPLORING
recipe
run_recipe
Do it again
recipe
✓
FAST
Your Chrome. Your logins.
Playwright scripts you own.
↺
Site changed? It repairs itself.
Get Started.
1
Set up gcontext
One-time setup. Already have a gcontext instance? Skip to step 2 ↓
Requirements: macOS, Linux, or Windows. Python 3.11+. Google Chrome. An MCP client (Claude Code, Cursor, Codex). No playwright install step: the agent drives your own Chrome.