Give an AI agent the page, not the cookie banner.
Framejet returns the rendered image inline through REST or MCP. Clean mode removes the overlays that otherwise become the model's entire visual context.
The problem
A vision model can only reason about the pixels it receives. If a consent wall covers the article, the model sees a consent wall — and a DOM dump cannot tell it whether the layout is broken.
How Framejet handles it
Connect the remote MCP endpoint once, or call the REST endpoint from an agent loop. When the useful page sits behind several clicks, pass a goal in plain words and exact values for anything that may be typed.
screenshot({
url: "https://en.wikipedia.org/",
goal: "Search for the Colosseum article. Stop when it is open.",
values: ["Colosseum"]
})Inline image
The MCP tool returns the PNG or JPEG directly to the agent, not a public URL it must fetch later.
No invented input
Goal mode may type only strings supplied in values. If none fits, it refuses the capture.
One account
REST and MCP authenticate with the same key and draw from the same monthly quota.