DOMShell — Browse the Web with Filesystem Commands (MCP Server)

[apireno]

Pre-submission Checklist

• I have verified that this discussion would not be more appropriate as an issue in a specific repository
• I have searched existing discussions to avoid duplicates

Discussion Topic

I built an MCP server that maps Chrome's Accessibility Tree to a virtual filesystem. Instead of screenshots and pixel coordinates, agents use ls, cd, grep, click, and type to navigate pages — the same way you'd work in a terminal.

GitHub: github.com/apireno/DOMShell
npm: npx @apireno/domshell

Why the Accessibility Tree?

Most browser automation feeds agents raw HTML or screenshots. The model burns through tool calls just figuring out what's on the page. Chrome's AX tree already solves this — it's a structured, role-annotated representation of the DOM that strips out layout noise and keeps semantics. DOMShell flattens it aggressively and maps it to a filesystem metaphor so agents can scope their work the way you'd cd into a directory.

In controlled testing (Claude, 4 web tasks, 8 trials), this cut average API calls per task from 8.6 to 4.3 compared to screenshot-based browsing. Full experiment data →

What It Looks Like

Agent: "Find the references on this Wikipedia article and extract the links"

1. ls                          → page sections (navigation/, main/, complementary/)
2. cd main/article             → scope to article content
3. grep -r references          → find the references section
4. cd references               → enter it
5. find --type link --meta     → all links with URLs in one call

Five calls. No screenshots. No coordinate math.

Quick Start

1. Install the Chrome Extension
2. Add to Claude Desktop config:

{
  "mcpServers": {
    "domshell": {
      "command": "npx",
      "args": ["-y", "@apireno/domshell", "--allow-write"]
    }
  }
}

38 tools across three security tiers (read-only by default, write requires --allow-write). Architecture and full tool list in the README.

What I'm Working On

Currently testing with smaller local models (Qwen3-4B, Llama3.2-3B) to see how the filesystem metaphor holds up with tighter context windows. Also exploring headless mode for CI pipelines. Would love feedback on the tool design or if anyone's tried a similar AX-tree approach.