GazeSight

GazeSight lets your coding agent see what you see, directly in the Chrome tab you're already working in. Browser vision for coding agents. Let your coding agent see what you see. **How it works** GazeSight connects your browser to a local MCP (Model Context Protocol) runtime that runs on your own machine. Your coding agent (Claude Code, Codex, or any MCP-compatible client) talks to that local runtime, which talks to the GazeSight extension over a local WebSocket connection. Nothing is captured until you explicitly authorize a tab with "Use this tab" — GazeSight never captures a page you haven't selected. **What it gives your agent** - Screenshots: viewport, full-page, a specific region, or a single element - Live DOM inspection: structure, attributes, computed layout - Console and network diagnostics for debugging - Safe, bounded interactions: click, type, select, scroll, hover — never arbitrary JavaScript execution - Visual diffing between two captures - A responsive preview window to check other viewport sizes without resizing your own window **Built local-first** - No GazeSight cloud capture backend. Captures go from your browser straight to your local MCP runtime. - Explicit, per-tab authorization — one click ("Use this tab") to opt a tab in, revocable at any time. - Least-privilege permissions: most capabilities work with only `activeTab`, `tabs`, `storage`, `scripting`, `webRequest`, and `alarms`. Background Capture (`tabCapture`/`offscreen`) and Continuous Access (`<all_urls>`) are optional, off by default, and requested only when you turn them on. - Manifest V3, no remote code, no `eval`. **Who it's for** Developers pairing a coding agent with real browser feedback — visually verifying UI changes, debugging layout issues, or letting an agent iterate on a page it can actually see. GazeSight is proprietary software. Website: https://www.gazesight.dev/ Support: support@gazesight.dev