Mobius

Let your AI coding agent drive and debug your app in your real, logged-in browser and see what it causes.

Mobius connects your browser to your AI coding agent. If you use Claude Code, Codex CLI, Gemini CLI, Cursor, or any other MCP compatible agent to build web apps, you've probably pasted console errors or network responses into chat just so the agent could see what's happening. Or typed out which buttons to click so it could reproduce a bug. Mobius skips both. Install the MCP, then click the toolbar icon to enable capture on a tab. From there your agent can see: * Console logs, warnings, and errors * Network requests and responses, including bodies * Page navigation, including SPA route changes * DOM changes, screenshots, and the live DOM/accessibility tree It can also act on what it sees. Your agent clicks, types, hovers, and navigates with real trusted browser input, the way a person would, not a scripted element.click() that misses half the page's event handlers. It finds what to click from a plain language description like "the newsletter signup field" instead of a CSS selector. It can reload the tab, run JavaScript in the page, or profile CPU and memory, similar to what you'd do by hand in Chrome DevTools. A cursor and a small on page HUD show what it's doing as it works, so you can watch instead of wondering. Usecases Debug a broken flow end to end. Tell the agent "signup is failing" and it drives to the failure itself, then comes back with the request that failed, the response body, and the console error that followed. Catch failures that look like successes. A 200 OK carrying {"success": false}. A click swallowed by an invisible modal. An auth header that quietly stops being sent partway through a flow. None of that shows up in a status code or an element.click(), which is exactly why it's easy to miss. Run a QA walkthrough before a release. Drive your key flows in one sequence and get a screenshot, a HAR, and the console output for each step, without maintaining a separate test suite. Turn a bug into something you can file. A confirmed repro becomes an ordered timeline, screenshots, and a HAR with full request and response bodies attached. Profiling; tell a stalled network apart from a blocked main thread apart from a memory leak, using request timing next to real CPU and memory profiles from your actual logged in session. See what a page is really sending. Every third party request, and what's in it: analytics, pixels, embedded widgets, and whether anything sensitive is leaking out in a request body. Since it drives the browser you're already signed into, this all works behind a login too: internal dashboards, admin panels, staging environments. No credentials handed to the agent, no separate test account. Why not just use an automated browser testing tool? Most AI browser tools drive a headless, separate browser. Mobius works on the tab you already have open. Click around, fill out a form, or trigger a bug yourself while your agent watches the same events and helps you fix them on the spot, or hand it the wheel and watch what it does. Everything stays local. Mobius only talks to a small MCP server on your own machine (mobius-mcp, open source, on npm). Nothing goes to a cloud service, and nothing runs until you turn it on for a tab. No account, no telemetry, no ads. Getting started 1. Install the mobius-mcp server. For Claude Code: $ claude mcp add mobius-mcp -- npx -y mobius-mcp Or add it to your MCP client's config directly. (Claude Code, Codex CLI, Cursor, and Gemini CLI etc) { "mcpServers": { "mobius-mcp": { "command": "npx", "args": ["-y", "mobius-mcp"] } } } 2. Click the Mobius icon and enable capture on the tab you're working on 3. Ask your agent to check the console, the network, the DOM, or just let it drive Repo: https://github.com/Topman-14/mobius-mcp