Review an unfamiliar site before an agent clicks
An agent click is a real visit. Read the host, what the click does, and any evidence on the page before you let the agent continue.
Drafted with AI tools (Claude Code and Codex) and published by Rattana Devs. How we write
Review an unfamiliar site before an agent clicks it. Read the host, decide what the click is allowed to do, and look at any evidence Kitsuvo shows. Then let the agent continue, open the page yourself, or stop. A filter score is an estimate of how the site was built, not a verdict that the page is safe to click.
We make Kitsuvo. This guide describes the checks a person can do in the browser before an assistant follows a link. It is not a promise that those checks find every scam.
Why does the click matter?
Before granting a read, check what stays on your machine when an agent reads a page: local browser data and copies received by your AI client have different lifecycles.
Search filtering does not visit results you leave closed. Kitsuvo does not open a result in the background to score it. When an agent clicks, the browser requests that URL. The site receives a normal visit. A form click can submit. A download click can save a file. On the normal signed-in profile, the visit can also send cookies that profile already has.
Page text can contain instructions aimed at the agent. Treat the page as data. The review is yours: read the address and the action before you accept the click. For what that visit stores, and what a connected client receives, see what stays on your machine when an agent reads a page.
What should you read in the address?
Look at the host, the part after the scheme and before the next slash. Link text, a button label and a page title can say something else. A look-alike name, an unexpected subdomain, or a host you do not recognize is a reason to stop and open the page yourself.
- Compare the host with the organization you meant to reach. A payment, account, or support page should sit on a host you already know, or on one that organization publishes.
- A builder hosting name, such as a
lovable.apporbolt.hostaddress, is a clue about how the site was generated. It is not, by itself, proof of a scam. The same is true of a custom domain with no builder name in the address. - Shortened and redirected links hide the host until they resolve. Open those yourself, or expand them, before an agent follows them.
The longer checklist for stores, certificates and contact details is in how to check if a website is legit. Look-alike sign-in pages are covered in how to spot phishing sites built with AI.
What is the click asking for?
Name the action before it runs. Opening a public article is a different decision from submitting a form, starting a download, approving a payment, or typing into an account page.
| What you see | What to decide first |
|---|---|
| A link to a public page | Whether that host is the one you meant |
| A form, checkout, or download | Whether you want that action, not only that page |
| A sign-in or password field | Leave it to yourself; do not hand the secret to the agent |
| Instructions in the page text | Ignore directions the page gives the agent until you have checked the host |
How should you read a Kitsuvo score?
If you open the page in Kitsuvo, the address bar can show a score from the evidence on that page. The bands are the same ones published with the product: below 40 is "No strong AI signs", 40 to 69 is "Possibly vibe coded", and 70 or above is "Likely AI-built". Each band is an estimate from builder traces and related signals. It is not proof of authorship, and it is not a safety certificate.
- A high score means Kitsuvo found signs associated with AI-builder construction or mass-produced pages. Many ordinary projects are built that way. Open the evidence list and decide whether this page is fit for the task.
- A low score means those signs were not found. The page can still be a scam, a copy, or simply unrelated to your task.
- Search results that match the filter fold into a "Hidden by Kitsuvo" line. You can open that line. Hiding a result is not the same as blocking the host. The difference between hiding images and blocking a site is a separate guide: hiding AI images versus blocking a site.
How the traces are collected is explained in how to tell if a website was made with AI.
What does Kitsuvo refuse by default?
On the current Mac release, local MCP can drive the browser only after you connect a client. Agent browsing is shown in a visible window with a Stop control, unless you specifically hide the window. Sign-in pages are not read or scripted, and password fields do not take typed text, unless you start Kitsuvo with --allow-sensitive. That flag widens access. Leave it off unless you intend the wider access.
The security page also lists sensitive categories the agent is refused from reading, scripting, screenshotting, or acting on, unless that same flag is set: webmail; banking, brokerage and crypto accounts; payments and checkout; password managers; cloud consoles; and GitHub, GitLab and Bitbucket settings. The agent may still open such a page and is told to hand the step to you. The lists are curated, not exhaustive. A page outside the list can still be the wrong place to type a secret.
For a site you do not trust, the security page recommends --separate-profile so the agent does not share the cookies of your normal signed-in profile. kitsuvo mcp uses that normal profile by default. The published Windows 0.2.4 installer cannot complete the current MCP connection check; see the Windows MCP compatibility limit. This guide's review steps still apply when you click yourself on that Windows build.
A review order that stays short
- Read the host aloud or copy it somewhere you can compare it.
- Say what the click will do: open, submit, download, pay, or sign in.
- If Kitsuvo has a score, open the evidence and treat it as a clue about construction, not as permission.
- If the page wants a password, a payment, or a mailbox, do that step yourself.
- If you are unsure, stop the agent and open the page in the window you are watching.
Stopping is a complete outcome. You do not owe the agent the click.
Questions people ask
Does a low Kitsuvo score mean the site is safe to click?
No. A low score means the page did not show strong signs of AI-builder construction. It says nothing definite about fraud, malware, or whether the host is the one you wanted.
Does a high score mean the agent should skip the site?
Not by itself. A high score means the evidence points to likely AI-built or mass-produced construction. Read the evidence, then decide if the page is fit for the task. You can open a held page anyway.
Will Kitsuvo click the link for me?
Only if you have connected an AI client and the client calls a browser tool. Ordinary filtering does not open results you did not choose. You can press Stop while an agent session is visible.
Should the agent type a password if the page looks right?
No. Sign-in pages and password fields stay protected unless you start Kitsuvo with --allow-sensitive. Even then, type secrets yourself. A matching name on the page is not a reason to hand a password to an agent.
Comments
Read the discussion or sign in to add a plain-text comment.
Loading comments…