Skip to content

Manual MCP configuration

If you installed the native Claude Code or Codex plugin, you do not need this page—the plugin already declares and manages Krometrail's MCP server.

Use manual configuration when you installed the standalone krometrail command or your MCP client cannot load the native plugin.

Add Krometrail from the command line

bash
claude mcp add --scope user krometrail -- krometrail mcp
codex mcp add krometrail -- krometrail mcp

Restart or reload the client, then check its MCP status for a connected Krometrail server.

Generic MCP client entry

json
{
  "mcpServers": {
    "krometrail": {
      "command": "krometrail",
      "args": ["mcp"]
    }
  }
}

The MCP client must be able to find krometrail on PATH. The default standalone install location is ~/.local/bin/krometrail.

Do not run mcp as a health check

krometrail mcp is a standard-input/output protocol server. It waits for an MCP client and writes protocol traffic—not a user interface—to standard output. Use these checks in a terminal instead:

bash
krometrail --version
krometrail doctor

Krometrail keeps protocol output isolated from diagnostics. Bounded private runtime logs live under the platform data directory; failed or degraded responses identify the relevant file and correlation identifier. Startup failures that occur before file logging is available fall back to standard error.

What the connection gives your agent

The server exposes browser lifecycle and page control, current screenshots and structured snapshots, temporal visual evidence, nearby browser events, and retained artifact/source-frame reads.

One MCP server owns at most one active browser session. Stop or detach from the current session before starting or attaching another one. Ending the MCP transport closes a Krometrail-managed browser or detaches from an externally owned browser.

Browser control and captured evidence stay local unless the connected agent explicitly reads a response or artifact through MCP.

For practical requests, see Use Krometrail with your agent. For startup failures, see Troubleshooting.