Kimi WebBridge Pro

The problem it addresses

Many browser automation tasks fail not because a click is impossible, but because the agent accidentally disturbs the user's browser state. If an agent treats every visible tab as disposable, it may close a logged-in user tab, retry a click before the page finishes loading, or miss a popup that was blocked or opened in the background.

Kimi WebBridge Pro is designed for this situation. It lets an AI agent control the user's live browser through a local daemon at http://127.0.0.1:10086, while keeping the difference between user-owned tabs and task-owned tabs explicit. Existing login state or the current tab can be attached with find_tab; a temporary tab that should be cleaned up later can be created with navigate plus newTab:true.

How it works

The workflow is intentionally narrow and operational:

• Check readiness without side effects: use doctor.py or the readiness report to verify binary presence, daemon status, port reachability, PID staleness, and extension connection.
• Assign one stable session name: this keeps the task's browser actions tied to a single context instead of drifting across multiple sessions.
• Read page state before acting: snapshot returns URL, title, accessible text, and @e references. For unknown pages, snapshot.py --auto is the default: small pages return compact output, while large snapshots are written to a UTF-8 JSON file.
• Use bounded actions: click and fill operate on snapshot references or selectors, while evaluate is reserved for limited page-state inspection or recovering a real link when normal actions are insufficient.
• Recover when the page does not change: after navigation or a click, use wait_for.py to check URL, title, or text conditions. If nothing changes, inspect list_tabs, look for popup blocking, and check whether the destination opened in a background tab.

This matters because browser pages are not always obvious state machines. Single-page apps can update in place, destination pages can open in new tabs, and result cards can contain nested click targets. The skill pushes the agent toward snapshot-first reasoning instead of blind DOM guessing.

Boundaries and cautions

It is not a general-purpose scraper, and it should not be used to bypass site security.

• Do not close user-owned tabs: tabs selected with find_tab belong to the user and should only be closed if the user explicitly asks.
• Do not inspect sensitive state: avoid returning cookies, authorization headers, session tokens, password fields, browser storage, or unrelated private page content.
• Do not bypass protections: do not attempt CAPTCHAs, paywalls, age gates, browser warnings, or other site security controls.
• Clean up temporary artifacts: screenshots, PDFs, and large snapshot files should be deleted after use unless the user asks to keep them.

It fits best when an agent needs to work with the user's current browser context for form filling, content extraction, page-state inspection, screenshot capture, or diagnosing why a click did not navigate as expected.