Symbolic Capital

Gorgon Eye

Chrome Extension AI Browser Copilot Manifest V3 Side Panel OpenClaw Integration

A Chrome extension that puts an AI agent in the browser sidebar. It reads the page you're on, understands context, takes actions โ€” clicking, typing, navigating, extracting data โ€” and remembers everything across sessions. It's not a chatbot overlay. It's a working copilot with eyes and hands.

Anthropic shipped "Claude in Chrome" months after this was already running. That product resets every session, can't touch local files, and is locked to one model. Gorgon Eye connects to the full OpenClaw agent fleet, maintains persistent memory, and works with whatever model is assigned to the browser agent role.

What It Does

Page Awareness Reads and understands the current page. Content injection captures DOM structure, text, forms, and interactive elements and feeds them to the agent as context.
Browser Actions Click elements, type text, fill forms, navigate pages, extract data. The agent sends structured commands that the content script executes.
Persistent Memory Everything the agent learns โ€” about your clients, your workflows, your preferences โ€” persists in memory files on your local filesystem. It knows you next session.
Agent Switching Not locked to one model. The side panel connects to any agent in the OpenClaw fleet. Switch between Iris (Sonnet), Medusa (Opus), or any other agent depending on the task.
File Operations Upload files, read local documents, write results back to the filesystem. The browser agent has the same workspace access as any other agent in the fleet.
Site Adapters Pluggable adapter architecture for different websites. Each adapter understands the specific DOM structure and interaction patterns of a particular site.

The Hard Parts

Manifest V3's security model fights you at every turn. Content scripts run in an isolated world. The page's actual JavaScript runs in the MAIN world. The background service worker has no DOM access. The side panel is its own context. Getting all four to communicate reliably โ€” especially when Chrome can suspend the service worker at any moment โ€” required a self-healing harness architecture that detects and recovers from broken message channels.

The biggest recurring failure mode: agents trying to type text into a field and hit Enter as two separate commands instead of using the atomic chat command. Under context pressure (long conversations, complex pages), agents regress to less efficient patterns. The fix was documentation in the agent's memory โ€” not code changes, but making sure the agent remembers the right way to do it.

CSS selector stability is another ongoing challenge. A critical fix โ€” div.font-claude-response โ€” was the difference between the extension working and silently failing on Claude's own interface. Site adapters exist to isolate these brittle selectors so the core logic doesn't break when a site updates their CSS classes.

Where It's Going

Workflow recording โ€” do a task once, the extension records the steps, replays them on demand. Scheduled tasks for recurring browser work. Inline rendering of visualizations and diagrams in the sidebar. These are on the roadmap, informed by what Anthropic shipped in their version and what's still missing.

The longer-term vision: this is one sensory channel for the agent fleet. Gorgon Eye is browser eyes. Serpent Wing is physical-world eyes (drone). SC-Mail is the inbox. The agents don't care which channel they're operating through โ€” the perception and action interfaces are interchangeable.

โ† Back to Projects