§ 03
Eyes and hands, over MCP
The server speaks MCP over stdio, so any MCP client can drive the same window you are looking at. Every tool takes a window argument, so a phone and an iPad can be driven side by side.
custom_ai_view_openopen a URL on a given device
custom_ai_view_waitwait for an element, or for loading to finish
custom_ai_view_screenshotthe live window — and it says so when it is not
custom_ai_view_findelements by selector or visible text, with real visibility
custom_ai_view_clicktap something
custom_ai_view_typetype into a field, optionally submitting on Enter
custom_ai_view_keyEnter, Escape, Tab, arrows
custom_ai_view_scrollthe page, or one scroller inside it
custom_ai_view_inspectmarkup, computed styles, box, ancestry
custom_ai_view_treewalk the DOM a level at a time
custom_ai_view_editchange the live page; survives reload; revert undoes it
custom_ai_view_consolepage output and the browser's own messages, with ages
custom_ai_view_backgo back without losing SPA state
custom_ai_view_recordscreen recording with no fixed length
custom_ai_view_collectscreenshot + markup + console + selection, in one folder
custom_ai_view_statewhat is open right now: address, device, viewport, selection
custom_ai_view_set_devicechange the device without losing the page
custom_ai_view_devicesthe whole catalogue, with points, pixels and millimetres
custom_ai_view_reloadnormal, hard, or clearing the cache first
custom_ai_view_editsevery live edit currently in force
custom_ai_view_revertput one of them back, or all of them
custom_ai_view_record_clipa clip of fixed length, when the length is known
custom_ai_view_librarywhat has been captured, filed by site
Add the server once. It starts the app itself if it is not running, so an agent never has to ask a human to open something first.
{
"mcpServers": {
"custom-ai-view": {
"command": "node",
"args": ["<path>/mcp/server.js"]
}
}
}
In VS Code, Custom AI View: Show MCP Server Setup hands you that block with the path already filled in.
Clients — the protocol is the contract, not the vendor
Claude Code
Claude Desktop
ChatGPT Desktop
Cursor
Windsurf
Zed
GitHub Copilot
Cline
Continue
Goose
…and anything else that speaks MCP
Nothing here is a special case. The server implements the Model Context
Protocol over stdio and declares its tools; whatever reads that declaration can drive the window. The
names above are the clients this has been run against — not a list of what is allowed.
Agent skill
It also ships as a skill
Tools tell an agent what it can call. They do not tell it when a screenshot is worth
taking, that the inspector beats guessing a selector, or that a device has to be set before a URL
is opened. skills/custom-ai-view/SKILL.md in the repository is that
second half — drop it in and the agent uses the tool the way someone who knows it would.
Read the skill