page_exec executes CDP-style commands inside the target page: { cmd: "DOM.querySelector", args: {...} }. The commands are implemented locally in the page context by the injected client (not native Chrome DevTools Protocol forwarding), with naming aligned to CDP style.
| Command | Description |
|---|
DOM.querySelector | Query an element by selector; returns a serialized node (tag/attrs/xpath/text/computed/framework) |
DOM.querySelectorAll | Batch query; returns a simplified node tree (excluding aiti's own UI, capped at 20 entries) |
DOM.getDocument | Document tree overview (configurable depth, default 3 levels) |
DOM.getBoxModel | Element box model (content quad + padding/margin/border/scroll) |
DOM.getOuterHTML | Outer HTML (truncated to 8000 characters) |
DOM.getInnerText | Visible text (sanitized, then truncated) |
DOM.setAttribute / DOM.removeAttribute | Attribute operations |
DOM.removeNode ⚠️ | Remove a node (irrecoverable, requires confirmation) |
DOM.clearSelection | Clear the current selection state and outline |
| Command | Description |
|---|
CSS.getComputedStyle | computed style |
CSS.getMatchedRules | Matched rules |
CSS.setStyle | Override styles |
CSS.listStyleSheets | Stylesheet list |
CSS.getStyleSheetText | Stylesheet text |
| Command | Description |
|---|
Input.click / Input.hover / Input.focus / Input.blur | Basic interactions |
Input.setValue | Set value (React-compatible: native setter + input/change events) |
Input.type / Input.insertText / Input.keyPress | Keyboard input |
| Command | Description |
|---|
Page.getMeta | Page metadata (title/url/viewport/framework detection, etc.) |
Page.scrollTo / Page.scrollIntoView | Scroll positioning |
Page.reload ⚠️ | Reload (requires confirmation) |
Page.navigate ⚠️ | Navigate (requires confirmation) |
Page.captureScreenshot 🖐 | Screenshot (getDisplayMedia; the browser pops a picker to choose screen/window/tab; large images are automatically written to disk and a path is returned) |
Page.getFrameTree | Frame tree (iframe src/name/title; same-origin includes url/readyState) |
| Command | Description |
|---|
Runtime.evaluate | Execute JS in the page context — disabled by default; enable it in the panel or call Runtime.setEvalEnabled {enabled:true} first; contextId enters an iframe context (same-origin, index starts at 1) |
Runtime.setEvalEnabled | Toggle evaluate |
Runtime.fetch | Send requests in the target page's context (same-origin, no CORS; affected by network emulation) |
Runtime.getProperties | Enumerate object properties |
| Command | Description |
|---|
Storage.localStorageGet / localStorageSet ⚠️ / localStorageClear ⚠️ | localStorage (writes/clears touch persistent data, require confirmation) |
Storage.sessionStorageGet | sessionStorage |
Storage.cookies | Cookies (document.cookie, truncated to 2000 characters) |
| Command | Description |
|---|
Performance.getEntries | Resource timing entries (most recent N, default 10) |
Operations that browser security policy requires transient activation for are executed on the page's behalf by the host page via the gesture bridge:
| Command | Description |
|---|
Browser.copy / Browser.paste | Clipboard write/read (reading may pop an authorization prompt) |
Browser.download | Download a file |
Browser.openWindow | Open a new window |
Browser.pickFile | File picker (small files return content directly / small images convert to dataUrl) |
Browser.share | System share panel (title/text/link) |
Browser.fullscreen | Fullscreen |
Browser.play | Play audio/video (bypasses autoplay policy) |
Legacy short names are auto-mapped via CMD_ALIASES: getInfo → DOM.querySelector, highlight → DOM.querySelector, scrollIntoView → Page.scrollIntoView. Use the standard names in new code.
| Concept | Description |
|---|
| Target | pageId (instanceId/frameId); default = host frame |
| Dangerous commands ⚠️ | DANGER_COMMANDS list (5 items) → confirmation card + confirmation queue; repeat confirmations can be waived within a debug session (dangerAutoOkMs) |
| Gesture bridge 🖐 | GESTURE_COMMANDS list (9 items) → executed by the host (transient activation-safe); unattended mode can be configured to auto-allow |
| Result | {ok: true, result} or {ok: false, error}, correlated by requestId |
For invoking via MCP, see MCP bridge · server.
