Flash AI Browser Assistant

Analyze, operate, debug, and summarize any web page with AI — bring your own model.

Pagkit AI Web Assistant is an AI browser-agent extension that works with your own models. Users chat with the AI in the browser side panel; the AI can read the current page and, with the user's per-site approval, perform clicks, typing, navigation, network inspection, console log reading, and JavaScript execution through the Chrome DevTools Protocol (CDP). Key features: Side panel chat — click the toolbar icon (or Ctrl/Cmd+Shift+K) to open it. Streaming answers, Markdown rendering, code highlighting and KaTeX; reasoning models show a collapsible thought process; conversation history is stored locally only. Bring your own model (BYOK) — any OpenAI-compatible endpoint. Presets for DeepSeek (default), Moonshot, OpenAI and local Ollama; custom providers can be added in the options page. API keys are stored only in local chrome.storage.local. Browser actions — the AI can read a page's main text, inspect its structure and take screenshots. Once approved, it can also click, type, scroll, choose options, press keys, open URLs and manage tabs. Developers can also have it inspect network requests, read console logs or run JavaScript. Safety & authorization — a permissions menu next to the composer offers two modes: "Requires approval" (default) allows page reads automatically and asks per site for navigation, click/type and JavaScript ("once" or "always", revocable in options); "Full access" runs everything without asking and marks the icon amber. A step limit and a stop button are included. Context — @ references open tabs, / runs skills and quick commands, and 📎 attaches txt/md/csv files plus text extraction from PDF and DOCX documents. Context menus — right-click on a page or a text selection to send it to the side panel for questioning. How to use: After installation, click the toolbar icon to open the side panel; a privacy notice appears on first use. Open the options page, pick a provider preset (DeepSeek by default), paste your API key, press "Test" to verify the connection, then "Fetch models" to load the model list, and finally "Save" to apply. Chat in the side panel. When asking the AI to operate pages (e.g. "open example.com and summarize it"), a per-site approval dialog appears; approve "once" or "always" and the tools run step by step. In the composer you can use @ to reference tabs, / to insert quick commands, and 📎 to attach files (PDF text is extracted). Press the stop button to abort a running task; the history panel reopens or deletes conversations; per-site approvals can be revoked under options → Site permissions.