MCP 服务器

Oxide

Intelligent LLM orchestrator that automatically routes tasks to the most appropriate AI model (Gemini, Qwen, Ollama, LM Studio) based on task characteristics, enabling distributed processing and parallel execution across local and network services.

README

Oxide - Intelligent LLM Orchestrator

Oxide is an intelligent orchestration system that allows Claude Code to automatically route tasks to the most appropriate LLM based on task characteristics, enabling distributed AI resource utilization across local and network services.

Features

Automatic Task Routing: Intelligently classifies tasks and routes them to the optimal LLM
Parallel Execution: Distributes large codebase analysis across multiple LLMs simultaneously
MCP Integration: Native integration with Claude Code via Model Context Protocol
Multi-Service Support: Works with Gemini CLI, Qwen CLI, Ollama (local & remote), and LM Studio
Web Dashboard: Real-time monitoring and configuration UI (coming soon)

Architecture

Claude Code (MCP) � Oxide Orchestrator � [Gemini | Qwen | Ollama | LM Studio]

Supported Services

Local Services:

Gemini CLI: Large context window (2M tokens) - ideal for codebase analysis
Qwen CLI: Code-specialized - best for code review and generation
Ollama: Local inference - fast, low-latency queries

Network Services (LAN):

LM Studio: OpenAI-compatible API on laptop
Ollama Remote: Distributed processing on server

Installation

# Clone the repository
cd /Users/yayoboy/Documents/GitHub/oxide

# Install dependencies
uv sync

# Verify installation
uv run oxide-mcp --help

Configuration

Configure services in config/default.yaml:

services:
  gemini:
    type: cli
    executable: gemini
    enabled: true

  qwen:
    type: cli
    executable: qwen
    enabled: true

  ollama_local:
    type: http
    base_url: "http://localhost:11434"
    enabled: true
    default_model: "qwen2.5-coder:7b"

Integration with Claude Code

Add to ~/.claude/settings.json:

{
  "mcpServers": {
    "oxide": {
      "command": "uv",
      "args": ["--directory", "/Users/yayoboy/Documents/GitHub/oxide", "run", "oxide-mcp"],
      "env": {
        "OXIDE_AUTO_START_WEB": "true"
      }
    }
  }
}

Note: Setting OXIDE_AUTO_START_WEB=true automatically starts the Web UI when the MCP server launches!

Quick Start

Launch All Services

Multiple ways to start Oxide:

# Option 1: Unified launcher (MCP + Web UI)
uv run oxide-all

# Option 2: Auto-start Web UI with MCP (set OXIDE_AUTO_START_WEB=true in settings.json)
uv run oxide-mcp

# Option 3: Shell script
./scripts/start_all.sh

# Option 4: Separate services
uv run oxide-mcp    # MCP server only
uv run oxide-web    # Web UI only

See AUTO_START_GUIDE.md for detailed auto-start configuration.

Usage

Once integrated with Claude, use the MCP tools:

# Intelligent task routing
Use oxide route_task to analyze this code for bugs

# Parallel codebase analysis
Use oxide analyze_parallel to analyze the ./src directory

# Check service status
Use oxide list_services to show available LLMs

Task Classification

Oxide automatically classifies tasks:

CODEBASE_ANALYSIS (>20 files or >500KB) � Gemini
CODE_REVIEW ("review" keyword) � Qwen
CODE_GENERATION ("generate"/"write" keywords) � Qwen/Ollama
QUICK_QUERY (simple, no files) � Ollama Local

Web Dashboard

Oxide includes a real-time web dashboard for monitoring and control:

# Start backend server
uv run oxide-web

# Start frontend (in another terminal)
cd oxide/web/frontend && npm install && npm run dev

Access at http://localhost:3000

Features:

Real-time service status monitoring
Task execution history
System metrics (CPU, memory)
WebSocket live updates
Service health checks

See WEB_UI_GUIDE.md for complete setup guide.

Network Services Setup

Configure remote LLM services on your LAN:

# Setup Ollama on another machine
./scripts/setup_ollama_remote.sh --ip 192.168.1.100

# Setup LM Studio on laptop
./scripts/setup_lmstudio.sh --ip 192.168.1.50

# Test network services
uv run python scripts/test_network.py --all

# Scan network for services
uv run python scripts/test_network.py --scan 192.168.1.0/24