Tool Modes

OpenForge MCP provides 3 tool modes that can be selected based on the AI client and model capabilities.

Mode Overview

Mode	Exposed Tools	Token Consumption	Recommended For
Full	All tools (100+)	High	Large models, debugging
Essential	Core tools only (~20)	Medium	Balanced usage
Dynamic (default)	3 meta-tools	Low	General use, local LLMs

Full Mode

Exposes all tools directly to the AI client. The AI can select and call tools straight from the list, making this the simplest mode of operation.

Advantages:

• Tool calls complete in a single step
• The AI knows all tools and parameters upfront

Disadvantages:

• Consumes a large number of tokens with every request
• Fills up the context window on local LLMs

Recommended when:

• API costs are not a concern
• All tools are used frequently during development

Essential Mode

Exposes only the most commonly used core tools. This includes tools for scene manipulation, object creation, material settings, screenshots, and other everyday tasks.

Advantages:

• Good balance between token consumption and convenience
• Covers most use cases

Disadvantages:

• Advanced tools (NavMesh, VFX, Terrain, etc.) cannot be called directly

Recommended when:

• Doing basic scene building or modeling
• You want to keep token consumption somewhat low

Dynamic Mode (Default)

Exposes only the 3 meta-tools (list_tools, get_tool_schema, call_tool). The AI fetches tool information as needed and calls tools dynamically.

Advantages:

• Minimal token consumption
• Context size stays the same regardless of tool count
• Runs comfortably on local LLMs

Disadvantages:

• Tool calls require 2-3 steps (list retrieval, schema retrieval, execution)
• The AI needs to understand how to use meta-tools

Recommended when:

• General use (default mode)
• Using a local LLM
• Optimizing API costs

Switching Modes

Via Environment Variable

# Dynamic mode (default)
export OPENFORGE_TOOL_MODE=dynamic

# Essential mode
export OPENFORGE_TOOL_MODE=essential

# Full mode
export OPENFORGE_TOOL_MODE=full

Via AI Client Configuration File

For Claude Desktop's claude_desktop_config.json:

{
  "mcpServers": {
    "openforge": {
      "command": "uvx",
      "args": ["openforge-mcp"],
      "env": {
        "OPENFORGE_TOOL_MODE": "essential"
      }
    }
  }
}

Via Startup Arguments

uvx openforge-mcp --tool-mode essential

Usage with Local LLMs

When using a local LLM, the context window is limited, so Dynamic mode is strongly recommended.

LM Studio Setup

1. Launch LM Studio and load an MCP-compatible model
2. Start the OpenForge MCP server in HTTP mode

uvx openforge-mcp --transport http --port 8080

3. Specify the server URL in LM Studio's MCP settings

http://localhost:8080

Ollama Setup

1. Start an MCP-compatible model with Ollama

ollama run qwen3

2. Connect to OpenForge MCP via an MCP client. Ollama alone does not natively support MCP, so use an MCP-compatible client application (e.g., Open WebUI) alongside it.

3. Start OpenForge MCP in HTTP mode

uvx openforge-mcp --transport http --port 8080

Tips for Local LLMs

• Use Dynamic mode to minimize token consumption
• With models that have small context windows, execute tool calls one at a time rather than chaining many calls at once
• The screenshot feature is particularly useful with multimodal models