LaunchRepo.

Glossary term

MCP server

A process that exposes tools to an AI agent through the Model Context Protocol — for LaunchRepo, the one that hands the agent a Chrome tab.

Definition

The Model Context Protocol (MCP) is an open standard for connecting AI agents to tools and data. An MCP server is a small program that lists what it can do — navigate, click, read a page, query a database — and the agent calls those tools as needed. Servers are registered in the agent’s configuration, typically as a command to start plus arguments.

Chrome DevTools MCP is the server that exposes a Chrome instance as tools. It is what makes a terminal agent a browser agent.

Sources: Model Context Protocol: What is MCP? · Model Context Protocol specification: Tools

Example

A developer in Newcastle uses Codex. He adds one MCP server entry to its configuration: npx starts chrome-devtools-mcp with --autoConnect. When the agent starts, the server lists its tools — open a page, click, fill a field, take a screenshot — and the agent calls them as it works. The generic entry from BROWSER-SETUP.md is below; the Codex page shows the same arguments in Codex’s own format.

A running server process is not proof yet. He asks the agent to open a public page and watches the tab in his own Chrome move. Only then does the first directory form begin.

{
  "mcpServers": {
    "chrome-devtools": {
      "command": "npx",
      "args": ["-y", "chrome-devtools-mcp@latest", "--autoConnect", "--no-usage-statistics", "--no-performance-crux"]
    }
  }
}

Why it matters for a directory launch

Without an MCP browser server a coding agent can only read and write files; it cannot fill a directory form. The server is also the boundary of trust: it controls the connected browser, so which Chrome it attaches to and how it is started decide what the agent can reach.

How LaunchRepo handles it

The repository documents a generic MCP server entry for chrome-devtools-mcp with --autoConnect (existing Chrome 144+) or --browserUrl (dedicated profile on a local port); where that entry goes depends on the agent — Claude Code, Codex, Cursor and six more each have their own configuration location, described on their pages here. Node and npm are needed for npx; a tested package version should be pinned after setup. A running server process alone proves nothing until a visible tab moves.

Go deeper

Connect your coding agent to your Chrome

Where the server entry goes, the two ways to attach it to Chrome and the test that proves the agent really controls your browser.

See the term in the repository

LaunchRepo is 299 directory playbooks, a product context file and a submission tracker your own agent runs. Pay €119.00 once; use it for every product you own.

See pricing