A real browser your AI can drive, with a person in control
CrawlGenie Browser is a browser for AI agents: real Chrome, run for your agents. Instead of pasting the whole page into the model, it hands over a tidy list of what is on the page and what can be clicked. Anything risky waits for a person. Every step is recorded so you can rewind.
Reads 3–8k tokens per look
The model sees a ranked list of the page, not 50–180k tokens of HTML. Re-checks return only what changed.
Risky clicks wait for you
Pay, send, delete and confirm buttons pause until a person approves, in the dashboard or on your phone.
Rewind and take over
Every step has a screenshot. Jump back, branch, or take the wheel live when a login or a CAPTCHA appears.
What happens when you use Browser?
Open a session
From the dashboard, ChatGPT, Claude or your code. Pick a saved login if the site needs one.
The agent works
It reads the tidy page view, clicks by reference, types with secrets it never sees, and reads only when it needs detail.
You stay in control
Approve risky actions, watch live, rewind any step, save the whole thing as a routine to replay without a model.
Verify it’s you
Enter the 6-digit code from your authenticator app.
{{totp:acme}} · the seed never leaves the server| Invoice | Date | Amount | Status | |
|---|---|---|---|---|
| INV-2291 | Oct 3 | $4,120.00 | Paid | Download PDF |
| INV-2297 | Oct 9 | $980.00 | Due | Download PDF |
| INV-2304 | Oct 14 | $2,375.00 | Due | Download PDF |
What do people use Browser for?
Supplier and customer portals
Export a report, file a form, check an order status, every night, with a saved login.
QA and operations
Walk through a flow, record it once, replay it with self-healing targets.
Research behind logins
Read dashboards and member areas your agent could not reach otherwise.
What you get, in detail
Why a dedicated browser for AI agents? Because the browsers people use were built to show pages, not to describe them. When an agent drives an ordinary browser, every step pastes 50,000 to 180,000 tokens of HTML or accessibility tree into the model, and the model spends most of its attention ignoring navigation, scripts and footers. CrawlGenie describes the page instead: a ranked list of what is visible and what can be clicked, each item with a stable reference, typically 3,000 to 8,000 tokens. After an action, only what changed comes back, often tens of tokens. The same session carries policy that the model cannot talk its way around: domain rules, a resource budget, and approvals that pause pay, send, delete and confirm actions until a person decides.
Logins are handled by profiles (cookies and storage saved under a name and locked while in use) and by the secrets vault, where a password is substituted on the server so the model only ever sees a placeholder. Every step is recorded with a screenshot and the storage state, so a run can be rewound, forked from any point, replayed as a routine without a model, or handed to a person live when a CAPTCHA or an unexpected screen appears. For developers the session exposes Playwright and Puppeteer over CDP, a tools API for any model, and an MCP server, and every snapshot carries its token count so the saving is visible per step.
For developers
Sessions expose a CDP endpoint for Playwright and Puppeteer, a tools API for any model, and an MCP server. API reference · Concepts.
What would this save you?
Move the sliders. The saving is computed from the measured ratios above and the model price you pick.
Assumes 10% of the page is kept per read (measured range 5–15%). Receipts in the product show the real number for your traffic.
Questions people ask
Does it solve CAPTCHAs?
No. It detects them, stops the loop so no tokens are wasted, and lets a person take over. Supporting sites can recognise the agent through its signed identity.
Can two agents share one login?
Only if you allow it; by default a saved login is used by one session at a time so cookies stay valid.
Is it a replacement for Playwright?
Playwright is the engine inside. CrawlGenie adds the model-friendly view, policy, approvals, recordings and the token accounting.
Spend tokens on answers, not on pages.
Create a workspace in a minute. Connect ChatGPT, Claude, Cursor or your own agents. Every answer comes with a token receipt.