An LLM agent living in the Chrome sidebar — give it tasks in plain language and it reads page snapshots, emits one JSON action per turn, and clicks and types for you, asking first before anything risky.

No new browser, no remote control — just an employee working in the Chrome sidebar. An MV3 extension built with WXT + React 19 + TypeScript, talking to OpenAI, DeepSeek, or any compatible endpoint, with keys stored locally only.

The agent loop runs “page snapshot → one JSON action → CDP execution”: text snapshots list only interactive elements, the model returns a single action per turn — slower, but every step is auditable. Input goes through real CDP events first, with DOM fallback marked in receipts. Component-library dropdowns (antd / element-plus) get a built-in two-step recipe, ask_user options render as buttons, and a 64px floating orb shows while it thinks.

High-risk actions (pay, delete, submit, order) are intercepted by a keyword regex and paused behind a confirmation card with three choices: deny, allow, or always allow for this session (in-memory, never persisted). E2E runs against a zero-dependency deterministic mock LLM state machine covering search, confirmation, canvas visual fallback, and format-tolerance scenarios — cheap and never moody.

Edg Agent