MCP and AI control

• search.web returns structured search results.
• web.read reads a URL as clean structured content.
• web.snapshot captures structured page state such as headings, links, forms, controls and tables.
• web.find and web.extract locate and extract specific information.
• readMany/searchAndRead support broader research flows.

MCP can use persistent authenticated Profiles when a task needs cookies, login state, JavaScript rendering, proxy/geographic context or other real browser state. This is the key difference between simple HTTP reading and acting through Social Browser.

• browser.tabs.screenshot captures managed tab content.
• browser.desktop.sources.list discovers monitors and application windows.
• browser.desktop.screenshot captures the primary monitor by default or a selected display/window.
• Desktop capture supports PNG/JPEG and crop regions.
• Screenshot tools return native MCP image content so the AI can see the image directly.

For new or unconfigured installations the MCP authorization preset defaults to full_automation. Explicit saved user choices such as balanced or cautious are preserved and are not overwritten by the product default.

Full Automation reduces approval friction for automation workflows, but it also allows host-sensitive operations such as desktop capture and process/filesystem actions. Use the preset that matches the trust level of the AI client connected to Social Browser.