Show HN: NavAgent – CDP-free, token-efficient web automation for AI
github.com
I use Claude Code daily, and I kept wanting it to interact with pages in my own browser, with my existing tabs, cookies, and logins.
Most browser automation approaches I tried for MCP were based on CDP / Playwright. I wanted something lighter and closer to normal browser behavior, so I tried a different design: the extension scans the DOM and sends the model a compact numbered list of actionable elements over a local WebSocket.
So the model sees something like: Search [input] Books [link] ... and can reply with browse_click(2) or browse_type(...)
That keeps scans very small (usually much smaller than screenshots or ARIA-tree dumps) and avoids screenshots, vision models, and huge accessibility payloads.
It currently handles SPAs, shadow DOM, same-origin iframes, and contenteditable editors. Everything runs locally. Setup is just the npm package plus the extension, and you can already try it by sideloading the extension from the repo and running npx navagent-mcp.
It's intentionally narrow: this is not a replacement for Playwright, CI automation or heavy scraping. It's a passive bridge for everyday browser tasks like opening pages, clicking around, and filling forms from an AI assistant.
It's still early, but it's already useful for my daily workflow. Feedback very welcome, especially on edge cases, security concerns and sites where it breaks.