DOMShell MCPサーバー

@apireno/domshell

MCP server that turns your browser into a filesystem. AI agents use ls, cd, grep, find, click, and type to browse the web — the same way you'd navigate a Linux filesystem.

DOMShell maps Chrome's Accessibility Tree to a virtual filesystem. Every DOM element becomes a file or directory. Agents work with familiar commands instead of raw selectors and coordinates.

Install

DOMShell supports three install paths — pick whichever fits your setup. Path 1 is the documented default and what 99% of users want; Paths 2 and 3 exist if you want lifecycle management across several MCP servers.

Path 1 — Native CLI (default, simplest)

npm install -g @apireno/domshell

Or run directly without installing:

npx @apireno/domshell

Works on macOS, Linux, and Windows with Node 18+. Nothing else required.

Path 2 — Dockerized (optional)

Run DOMShell in a container — useful if you don't want a global Node install or you want process isolation. Requires Docker Desktop. The token is persisted to a gitignored .env file so it survives container restarts and host reboots:

git clone https://github.com/apireno/DOMShell && cd DOMShell/mcp-server
cp .env.example .env
# Edit .env and set DOMSHELL_TOKEN to: $(openssl rand -hex 24)
docker compose build      # produces domshell-mcp-server:latest
docker compose up -d
docker compose logs -f

The container maps ports 3001 (MCP HTTP) and 9876 (WebSocket bridge to the Chrome extension) to your loopback interface, so the Chrome extension reaches it exactly as it would the native install. restart: unless-stopped brings the container back automatically when Docker Desktop starts (toggle "Start Docker Desktop when you log in" in Docker Settings to make Path 2 fully reboot-resilient).

Path 3 — ToolHive-managed (optional)

Use ToolHive (thv) to manage the container's lifecycle alongside any other MCP servers you're running. Same .env file Path 2 uses, just sourced by thv:

brew tap stacklok/tap && brew install thv      # one-time
cd DOMShell/mcp-server
cp .env.example .env       # edit DOMSHELL_TOKEN if you haven't already
docker compose build       # produces domshell-mcp-server:latest

thv run \
  --name domshell-mcp-server \
  --transport streamable-http \
  --target-port 3001 \
  -p 9876:9876 \
  --env-file .env \
  domshell-mcp-server:latest

thv list                          # workload status (URL on the proxy port)
thv logs domshell-mcp-server      # tail logs

The container is reachable on http://127.0.0.1:3001/mcp because --target-port 3001 auto-publishes that port to host loopback — your existing MCP client config (Claude Desktop, Cursor, …) keeps working without changes.

Path 3 needs a one-time launchd agent for reboot autostart (Path 2 survives reboot through Docker Desktop alone). Full reboot table, launchd plist template, autostart script, and a simulated-reboot verification: docs/deploy/container-and-toolhive.md.

Chrome extension (required for all three paths)

You also need the DOMShell Chrome Extension — the MCP server talks to the browser through it.

Quick Start

npx @apireno/domshell init

The setup wizard detects installed MCP clients (Claude Desktop, Cursor, Windsurf), generates a secure token, and writes the config. Use --yes for non-interactive mode.

Then:

1. Install the DOMShell Chrome Extension
2. Open Chrome's side panel and start a DOMShell session
3. Restart your MCP client — DOMShell tools will appear
4. In the DOMShell terminal, run connect <token> (printed by the wizard)

Claude Desktop Config

Add this to your Claude Desktop MCP settings (claude_desktop_config.json):

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

For the stdio proxy (required if your client needs command/args format):

{
  "mcpServers": {
    "domshell": {
      "command": "npx",
      "args": ["-y", "-p", "@apireno/domshell", "domshell-proxy", "--port", "3001", "--token", "YOUR_TOKEN"]
    }
  }
}

Benchmarks

We tested DOMShell against Computer-in-the-Cloud (CiC) — both using Claude as the underlying model — across 4 web tasks over 8 trials.

Metric	DOMShell	CiC
Avg API calls per task	4.3	8.6
Hardest task (T4)	6.0 calls	13.0 calls
Cold start vs CiC warm cache	4.5 calls	5.5 calls

DOMShell uses 2× fewer API calls to complete the same tasks. The filesystem metaphor gives the model a mental map of the page, so it spends less time exploring and more time extracting.

Full experiment data: experiments/claude_domshell_vs_cic

MCP Interface

Default (recommended): the single domshell_execute tool. Pass any DOMShell command as a string — "ls", "cd tabs/123", "open https://example.com" — or pass several newline-separated for a whole workflow in one call. One tool, one approval, the full command vocabulary in the tool description.

Multi-line semantics: when command contains newlines, each line runs in order in the same MCP session and lane, so cwd, env, and history persist between lines. An error on one line does not halt the rest — its message is included in the combined output and subsequent lines still run. This is the right shape for cleanup-line idioms like "cd path\ngrep pattern\ncd back": the trailing restore runs even if the middle step errors. Implementation: mcp-server/index.ts:1115-1136.

Lanes & multi-agent. Every reply ends with [lane: <id>]. Multiple MCP clients can connect simultaneously, each in its own isolated Chrome tab group. To carve sub-lanes within a single client (e.g. two Claude Desktop chats), pass group_id: "new" on the first call and carry the returned id thereafter. See CHANGELOG.md.

Granular mode — --granular

Start the server with --granular and the 38 per-command tools are exposed alongside domshell_execute. Use it when you want per-operation approval in the client UI.

Read tier (always available): ls, cd, pwd, cat, find, grep, tree, text, read, tabs, here, refresh, diff, eval, functions, watch, for, script, each, extract_links, extract_table

Write tier (--allow-write): click, focus, type, scroll, navigate, open, submit, back, forward, close, select, js, screenshot, wait, call

Sensitive tier (--allow-sensitive): whoami

CLI Flags

Flag	Description
--allow-write	Enable write-tier tools (click, type, navigate, etc.)
--allow-sensitive	Enable sensitive-tier tools (whoami)
--allow-all	Enable all tiers
--granular	Expose the 38 per-command tools alongside domshell_execute
--port N	WebSocket port (default: 9876)
--mcp-port N	HTTP MCP port (default: 3001)
--domains a.com,b.com	Restrict to specific domains
--token TOKEN	Set auth token (auto-generated if omitted)
--log-file PATH	Audit log location (default: audit.log)
--confirm	Opt in to per-action y/n prompts in the server terminal before each write. Off by default.
--no-confirm	No-op (kept for backward compatibility — per-action prompts are off by default).

Security

Every command goes through a 4-tier security model:

• Read — always allowed (ls, find, grep, text)
• Navigate — requires --allow-write (navigate, open, back, forward)
• Write — requires --allow-write (click, type, js, select)
• Sensitive — requires --allow-sensitive (whoami, cookie access)

All commands are logged to an audit file. Domain allowlists restrict which sites the agent can access. Auth tokens protect the HTTP endpoint.

Architecture

MCP Client (Claude, Cursor, etc.)
    ↓ HTTP :3001/mcp
DOMShell MCP Server (Express + WebSocket)
    ↓ WebSocket :9876
DOMShell Chrome Extension (CDP 1.3)
    ↓ Chrome Debugger Protocol
Browser DOM + Accessibility Tree

Links

• Chrome Web Store: pireno.com/domshell
• GitHub: github.com/apireno/DOMShell
• Blog: Why I Built a Filesystem for the Browser
• Project home: pireno.com/domshell

Built by Pireno.

License

MIT