What the Agent Can Do
Through the Browser Relay extension, the agent gets a discrete set of operations against your connected Chrome. This page is the catalogue.
Tab management
| Method | What it does |
|---|---|
browser.tabs.list | Enumerate every open tab with URL, title, active state, window ID, audible flag, load status |
browser.tabs.create | Open a new tab, optionally with a URL |
browser.tabs.navigate | Load a URL in an existing tab |
browser.tabs.reload | Refresh a tab |
browser.tabs.close | Close a tab |
These are stateless from the extension's side — every call goes through Chrome's standard chrome.tabs API.
Screenshots
| Method | What it does |
|---|---|
browser.tabs.capture | Capture the visible region of a tab as JPEG (quality 80) base64, with viewport width and height. Uses CDP Page.captureScreenshot; falls back to chrome.tabs.captureVisibleTab if CDP is unavailable. |
Returned images include the dimensions so the agent can compute pixel-accurate coordinates if it later wants to act on what it saw.
Debugger Protocol (CDP)
| Method | What it does |
|---|---|
browser.debugger.attach | Attach the Chrome debugger to a tab (protocol version 1.3) |
browser.debugger.detach | Detach the debugger |
browser.debugger.sendCommand | Forward an arbitrary CDP method (e.g. Runtime.evaluate, Page.captureScreenshot, DOM.describeNode) |
CDP is the most powerful surface — through sendCommand the agent can:
- Evaluate JavaScript in page context (
Runtime.evaluate). - Walk the DOM (
DOM.getDocument,DOM.querySelector). - Listen to events — JS dialog opens, network activity, page lifecycle.
The extension auto-dismisses JS dialogs (alert / confirm) when the debugger is attached, so they don't block automation.
DOM helpers
The desktop client wraps common CDP idioms into higher-level helpers the agent calls directly:
| Operation | What it does |
|---|---|
| Read page content | Title, URL, full innerText (capped at 60,000 chars, with a truncation flag) |
| Summarise DOM | Extract up to 200 interactive elements (buttons, links, inputs, custom roles) with CSS selectors, bounding boxes, visibility, text/aria-label, disabled state |
| Click by selector | Click an element identified by CSS selector |
| Type by selector | Type into an input identified by CSS selector |
| Scroll | Scroll the page or a specific scroll container |
| Upload file | Trigger a file input by CSS selector with a local file path |
The agent typically:
- Captures a screenshot to see the page.
- Summarises the DOM to learn the selectors.
- Calls click / type with specific selectors.
What the agent cannot do through the relay
- Persist state across browser restarts. The extension itself stores nothing about your browsing.
- Read your browser history or bookmarks. Those Chrome APIs are not in the extension's permission set.
- Modify Chrome settings. The extension does not request
chrome.managementor similar. - Touch other extensions. The relay has no inter-extension messaging.
- Operate in incognito mode. Manifest V3 + the
debuggerpermission do not extend to incognito by default. If you want incognito support, enable it explicitly per extension inchrome://extensions; otherwise incognito tabs are invisible to the agent.
Concurrency
The relay attaches the debugger to one tab at a time per operation. The agent can drive multiple tabs but does so sequentially. A "follow this link, screenshot the page, come back, click another" workflow is normal; a "do these three things in parallel" workflow is not.
Limits and quotas
- Screenshot quality is fixed at JPEG 80 — high enough for vision-capable models, small enough to fit in context.
- innerText is capped at 60,000 chars. Anything longer is truncated with an explicit flag the agent can see.
- DOM summarisation is capped at 200 interactive elements per call. For pages with more, the agent paginates by scrolling and re-summarising.