Oxide

Oxide is an intelligent MCP-backed orchestrator that automatically routes tasks to the most suitable LLMs, enabling distributed AI work across local and network services. It integrates with Claude Code to split, assign, and manage workloads efficiently, making large analyses and code projects faster and more scalable.

How to use

You interact with Oxide through the MCP client to route tasks, run parallel analyses, and check service status. Start the MCP server, connect your client, and use the built‑in tools to route code analysis, perform parallel reviews, and monitor task progress in real time.

How to install

Prerequisites you need before installing: Python 3.11+, uv package manager, and the necessary CLIs for local and network services. Follow the concrete steps below to set up Oxide on your machine.

# 1) Clone the Oxide repository
cd /Users/yayoboy/Documents/GitHub/oxide

# 2) Install dependencies (using the UV toolkit)
uv sync

# 3) Verify installation and available MCP commands
uv run oxide-mcp --help

Configuration

Configure local and network LLM services to be used by Oxide. The following examples show how to enable Gemini, Qwen, and Ollama for local use, and how to point Ollama Remote or LM Studio on the LAN.

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"

Web dashboard and monitoring

Oxide provides a real‑time web dashboard to monitor services, task history, and system metrics. You can start the web backend and then run the frontend to interact with the dashboard.

# Start backend server
uv run oxide-web

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

# Access the dashboard at
http://localhost:3000

Network services setup

For networked LLM services, configure remote instances on your LAN. You can set up Ollama Remote and LM Studio on separate machines and test connectivity from the local controller.

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

Notes and troubleshooting

Key tips to keep in mind: ensure the web UI starts when the MCP server launches if you want a dashboard, and verify that local services are reachable at their base URLs before starting routing tasks. If you modify the service configuration, reload or restart the MCP server to apply changes.

Available tools

route_task

intelligently routes a given task to the most suitable LLM based on task characteristics to maximize quality and speed.

analyze_parallel

performs parallel analysis of a codebase across multiple LLMs to accelerate large-scale reviews.

list_services

queries and displays the status of available LLM services and their health.

start_dashboard

launches the web dashboard for real-time monitoring and control.

test_network

provides utilities to test connectivity to networked MCP services and scan LAN visibility.