MCP Server Plugins

Plugins can expose MCP servers to Claude Code by including a .mcp.json file in the plugin root. Athena merges all loaded plugins' MCP configs and passes the result to the spawned Claude process via --mcp-config.

How MCP Integration Works

1. Each plugin directory that contains a .mcp.json file contributes its mcpServers entries to a merged config.
2. Athena writes the merged config to a temp file at startup: /tmp/athena-mcp-<PID>.json.
3. This file is passed to Claude Code via the --mcp-config flag.
4. Claude Code manages the MCP server processes; Athena does not start or monitor them directly.

MCP server names must be unique across all loaded plugins. A collision causes a startup error:

MCP server name collision: "my-server" is defined by multiple plugins.
Each MCP server must have a unique name across all plugins.

Plugin .mcp.json Format

{
  "mcpServers": {
    "my-server": {
      "command": "node",
      "args": ["./server/index.js"],
      "env": {
        "MY_VAR": "value"
      }
    }
  }
}

The format follows Claude Code's standard MCP config format. The command and args fields specify how to start the server process.

Skills with MCP Access

When a skill in a plugin that has a .mcp.json is invoked, it runs with the plugin's MCP server's tools available to Claude. This is how the e2e-test-builder plugin's browser automation skills get access to the agent-web-interface tools.

agent-web-interface

agent-web-interface is a browser automation MCP server built specifically for LLM agents. It produces semantic page snapshots instead of raw DOM dumps, optimized for token efficiency and reliable agent-driven navigation.

The Problem It Solves

Most browser automation tools expose entire DOMs or accessibility trees to the model, leading to rapid token exhaustion. agent-web-interface presents pages in a form aligned with how language models reason about interfaces — compact, structured, focused on user-visible intent.

Benchmarks

Benchmarks against Playwright MCP show approximately 19% fewer tokens consumed and approximately 33% faster task completion on common navigation tasks. Results are task-dependent.

How It Works

1. The browser is controlled via Puppeteer and Chrome DevTools Protocol (CDP)
2. Pages are reduced into semantic regions and actionable elements
3. A structured snapshot is generated and sent to the agent
4. Actions are resolved against stable semantic identifiers rather than fragile CSS selectors

CLI Arguments

The browser launches automatically on the first tool call — no explicit initialization step needed.

Environment Variables

Connecting to Your Chrome Profile (Chrome 144+)

To connect with your existing bookmarks, extensions, and logged-in sessions:

1. Navigate to chrome://inspect/#remote-debugging in Chrome
2. Enable remote debugging and allow the connection
3. Start the server with --autoConnect

{
  "mcpServers": {
    "agent-web-interface": {
      "command": "npx",
      "args": ["agent-web-interface@latest", "--autoConnect"]
    }
  }
}