This guide explains how RetroRecon's embedded MCP server works. The application now uses the RetroReconMCPServer class directly so no separate mcp-server-sqlite process is launched.

1. Run launch_app.sh (Linux/macOS) or launch_app.bat (Windows) to install the Python dependencies.
2. The Flask app automatically instantiates RetroReconMCPServer whenever a database is created or loaded. No separate background process is required.
3. There is no need to configure a port or connect LM Studio; the server runs entirely inside the application.
4. Set RETRORECON_LOG_LEVEL=DEBUG before launching to view MCP debug output in the console. Debug logs now include detailed information on which MCP servers are mounted, any health check failures, and when alternate API bases are used for fallback requests.

The embedded server now aggregates tools from all configured MCP services at runtime. When a new service is mounted (locally or remotely), its tool list is queried during startup. Healthy services are mounted under their configuration name and their tools automatically become available to the LLM with the prefix <name>_. No manual code changes are required for additional modules—simply add an entry in the MCP configuration.

External servers are defined in a single JSON or YAML file such as mcp_servers.json. Set the path with the RETRORECON_MCP_SERVERS_FILE environment variable if needed. Each entry specifies how the server should be launched or contacted:

- name: memory
  transport: stdio
  command: ["basic-memory", "mcp", "--transport", "stdio"]
  model: memory
  description: A memory server for storing and retrieving information.

Place this file where your RetroRecon instance can load it (for example in the project root) and ensure the application reads it on startup.

When using the stdio transport, RetroRecon starts the MCP server as a subprocess. A simplified example looks like this:

import subprocess

proc = subprocess.Popen(
    ["basic-memory", "mcp", "--transport", "stdio"],
    stdin=subprocess.PIPE,
    stdout=subprocess.PIPE,
    text=True,
)

RetroRecon keeps the process handle and communicates with the server through the captured stdin and stdout streams.

To query a running server for its tool metadata, create a FastMCP client. The client can introspect the server and return its registered tools:

from fastmcp import StdioClient

client = StdioClient(proc.stdout, proc.stdin)
metadata = client.get_metadata()
print(metadata.tools)

The returned list contains tool names and schemas, which RetroRecon uses to register each tool under the <server>_ prefix for the LLM.