CLI Mode (MCP Relay)

When LocalGPT Gen runs interactively with a CLI backend — Claude CLI, Gemini CLI, or Codex — it uses a relay architecture to ensure tool calls go to the same Bevy window you're looking at.

The Problem

Without the relay, each CLI backend spawns its own localgpt-gen mcp-server process. This creates a second Bevy window that opens, renders the world, and immediately closes — while the original window stays empty.

  ┌──────────────┐     spawns      ┌──────────────────┐
  │  Claude CLI  │────────────────►│ localgpt-gen      │ ← second window!
  │  (backend)   │   MCP stdio     │ mcp-server        │
  └──────────────┘                 │ (NEW Bevy + scene) │
                                   └──────────────────┘

  ┌──────────────────────────────────┐
  │  localgpt-gen (interactive)       │ ← your window (empty)
  │  Bevy window you're looking at    │
  └──────────────────────────────────┘

The Fix: MCP Relay

LocalGPT Gen starts a TCP relay server alongside the Bevy window. When a CLI backend spawns localgpt-gen mcp-server --connect, it connects to the relay instead of creating its own Bevy instance. All tool calls route to your existing window.

  ┌──────────────────────────────────────────────────────────────────┐
  │  localgpt-gen (interactive)                                       │
  │                                                                   │
  │  ┌────────────┐   GenBridge    ┌────────┐    TCP :9878            │
  │  │ Bevy 3D    │◄─────────────►│ MCP     │◄──────────────┐        │
  │  │ Window     │  (in-process)  │ Relay   │               │        │
  │  │            │                │ Server  │               │        │
  │  └────────────┘                └────────┘               │        │
  │        ▲                                                │        │
  │        │ renders scene                                  │        │
  └────────┼────────────────────────────────────────────────┼────────┘
           │                                                │
           │                          ┌─────────────────────┤
           │                          │  stdio ↔ TCP bridge │
           │                          │                     │
           │                     ┌────┴──────────────┐      │
           │                     │ localgpt-gen       │      │
           │                     │ mcp-server         │◄─────┘
           │                     │ --connect           │  TCP relay
           │                     └────┬──────────────┘
           │                          │ MCP stdio
           │                     ┌────┴──────────────┐
           │                     │  Claude CLI /      │
      you see the                │  Gemini CLI /      │
      world here                 │  Codex             │
                                 └───────────────────┘

Key points:

• Only one Bevy window exists — the one you see
• The relay server listens on TCP port 9878
• --connect makes the MCP server process a thin stdio↔TCP bridge
• Tool calls travel: CLI → MCP stdio → bridge → TCP → relay → GenBridge → Bevy
• Results travel the reverse path back to the CLI

How It Works

When you run localgpt-gen interactively:

1. Bevy window opens on the main thread
2. MCP relay starts on port 9878 (background thread)
3. Agent loop starts — you see:

   MCP relay active on port 9878 (CLI backends will use this window)
   CLI backend detected (claude-cli/opus). Gen tools will route to this window via MCP relay.

4. When the agent sends a prompt to the CLI backend:
   • The CLI backend reads its MCP config
   • It spawns localgpt-gen mcp-server --connect
   • The --connect process connects to the relay on port 9878
   • Tool calls flow through to the existing Bevy window
5. You see the world being built in real-time in your window

Configuration

Automatic (recommended)

If you run localgpt-gen interactively with a CLI backend model, it automatically configures the relay. No manual setup needed — just:

# Set your model to a CLI backend
localgpt-gen   # uses default model from config.toml

If config.toml has default_model = "claude-cli/opus" (or any gemini-cli/* / codex-cli/* model), the relay activates automatically.

Claude CLI

Add to project-level .mcp.json:

{
  "mcpServers": {
    "localgpt-gen": {
      "command": "localgpt-gen",
      "args": ["mcp-server", "--connect"]
    }
  }
}

Gemini CLI

Add to ~/.gemini/settings.json or workspace .gemini/settings.json:

{
  "mcpServers": {
    "localgpt-gen": {
      "command": "localgpt-gen",
      "args": ["mcp-server", "--connect"]
    }
  }
}

Codex CLI

Add to ~/.codex/config.json:

{
  "mcpServers": {
    "localgpt-gen": {
      "command": "localgpt-gen",
      "args": ["mcp-server", "--connect"]
    }
  }
}

These config files live in your LocalGPT workspace (default: ~/.local/share/localgpt/workspace/). They're created automatically on first run.

Migrating from standalone MCP

If you previously configured localgpt-gen mcp-server without --connect, update your configs to add "--connect" to the args array. This prevents the duplicate window issue.

Standalone MCP mode

If you want to run gen as a standalone MCP server (no interactive mode, the CLI is the sole orchestrator), use mcp-server without --connect:

localgpt-gen mcp-server

This is for desktop apps (Claude Desktop, Codex Desktop) and editors (VS Code, Zed, Cursor) where you don't need the interactive terminal. See MCP Server for details.

CLI Flags

localgpt-gen mcp-server --connect [PORT]

Flag	Description
--connect	Connect to an existing gen process's MCP relay
--connect 9878	Connect to a specific port (default: auto-discover)
--headless	Run without a window (for CI/batch generation)

Port auto-discovery reads from $XDG_RUNTIME_DIR/localgpt/gen-mcp-relay.port, which the interactive gen process writes on startup.

Supported Backends

Backend	Model prefix	MCP config location
Claude CLI	claude-cli/*	.mcp.json in workspace
Gemini CLI	gemini-cli/*	.gemini/settings.json in workspace
Codex CLI	codex-cli/*	.mcp.json in workspace

Troubleshooting

"No relay port specified and no running gen process found"

The --connect process can't find the relay. Make sure localgpt-gen is running interactively first:

# Terminal 1: start gen interactively
localgpt-gen

# Terminal 2: the MCP connect process finds it automatically
localgpt-gen mcp-server --connect

Still seeing two windows

Check that your MCP configs use --connect:

# Check workspace configs
cat ~/.local/share/localgpt/workspace/.mcp.json
cat ~/.local/share/localgpt/workspace/.gemini/settings.json

Both should have ["mcp-server", "--connect"] in the args. If they show ["mcp-server"] without --connect, update them.

Relay port conflict

If port 9878 is in use, the relay picks a random available port. Check the startup log for the actual port:

MCP relay active on port 51234 (CLI backends will use this window)

The --connect flag auto-discovers the port, so this usually isn't a problem.