OpenWebUI with Ollama, LiteLLM, MCPO, and MCP Gateway Deployment Guide¶

This guide provides a comprehensive walkthrough for deploying OpenWebUI with local LLM models via Ollama, integrated with MCP (Model Context Protocol) tools through MCPO and MCP Gateway.

🏗 Architecture Overview¶

This deployment creates a powerful AI assistant platform with local model hosting and extensible tool capabilities:

flowchart TD
  %% User Interface
  U[🌐 User Browser] -->|HTTPS :443| OW

  %% Core Components
  subgraph "OpenWebUI Deployment"
    OW[OpenWebUI Container 🖥️<br/>Web Interface]
    LL[LiteLLM Proxy 🔄<br/>Model Gateway<br/>Port 4000]
    DB[PostgreSQL Database 🗄️<br/>Port 5432]
  end

  %% Local AI Backend
  subgraph "Local AI Backend"
    OL[Ollama Server 🤖<br/>Port 11434<br/>Local LLM Models]
  end

  %% MCP Integration Layer
  subgraph "MCP Integration"
    MCPO[MCPO Server 🔌<br/>Port 8000<br/>MCP → OpenAPI Bridge]
    MCG[MCP Gateway 🚪<br/>Port 4444<br/>MCP Registry & Federation]
    MCP1[MCP Server 1 📦]
    MCP2[MCP Server 2 📦]
    MCPN[MCP Server N 📦]
  end

  %% External Services
  subgraph "External Services"
    OP[OAuth Provider 🔐<br/>Port 443]
  end

  %% Connections
  OW -.->|"OAuth Login"| OP
  OW -->|"Model API Calls"| LL
  OW -->|"Tool API Calls"| MCPO
  OW -->|"DB Connection"| DB

  LL -->|"Model Requests"| OL

  MCPO -->|"Register & Proxy"| MCG
  MCG -->|"MCP Protocol"| MCP1
  MCG -->|"MCP Protocol"| MCP2
  MCG -->|"MCP Protocol"| MCPN

  %% Styling
  classDef deployment fill:#1F618D,stroke:#85C1E9,stroke-width:2px,color:#FFF;
  classDef ai fill:#27AE60,stroke:#58D68D,stroke-width:2px,color:#FFF;
  classDef mcp fill:#8E44AD,stroke:#BB8FCE,stroke-width:2px,color:#FFF;
  classDef external fill:#E74C3C,stroke:#F5B7B1,stroke-width:2px,color:#FFF;

  class OW,LL,DB deployment;
  class OL ai;
  class MCPO,MCG,MCP1,MCP2,MCPN mcp;
  class OP external;

Components Overview¶

1. OpenWebUI: Feature-rich web interface for AI interactions
2. Ollama: Local LLM runtime supporting various models (Llama, Mistral, Granite, etc.)
3. LiteLLM: Unified proxy providing OpenAI-compatible API for multiple model backends
4. MCPO: Bridges MCP servers to OpenAPI, making tools accessible to OpenWebUI
5. MCP Gateway: Central registry and federation point for MCP servers
6. PostgreSQL: Persistent storage for OpenWebUI data

📋 Prerequisites¶

• Operating System: Linux, macOS, or Windows with WSL2
• Docker/Podman: Container runtime installed and running
• Python 3.8+: For running MCPO and MCP tools
• 8GB+ RAM: Recommended for running local models
• 20GB+ Storage: For model files and data

🚀 Deployment Steps¶

Step 1: PostgreSQL Database Setup¶

First, set up a PostgreSQL database for OpenWebUI's persistent storage:

# Create a Docker network for service communication
docker network create openwebui-net

# Run PostgreSQL container
docker run -d \
  --name postgres \
  --network openwebui-net \
  -e POSTGRES_DB=openwebui \
  -e POSTGRES_USER=openwebui \
  -e POSTGRES_PASSWORD=changeme \
  -v postgres_data:/var/lib/postgresql/data \
  -p 5432:5432 \
  postgres:15-alpine

# Verify PostgreSQL is running
docker logs postgres

Step 2: Install and Configure Ollama¶

2.1 Install Ollama¶

Important: Install Ollama on the host system, not in a container or WSL2.

# Universal installer for macOS/Linux
curl -fsSL https://ollama.com/install.sh | sh

# For Windows, download from: https://ollama.com/download

2.2 Pull AI Models¶

# Pull IBM Granite model (recommended)
ollama pull granite3.3

# Alternative smaller model for lower-end hardware
ollama pull granite3.3:2b

# Pull additional models as needed
ollama pull llama3.2
ollama pull mistral

# List installed models
ollama list

# Test model
ollama run granite3.3 "Hello, how are you?"

2.3 Configure Ollama for Network Access¶

By default, Ollama only listens on localhost. To make it accessible to containers:

# Linux/macOS: Edit systemd service
sudo systemctl edit ollama.service

# Add these lines:
[Service]
Environment="OLLAMA_HOST=0.0.0.0"

# Restart Ollama
sudo systemctl restart ollama

# Verify it's listening on all interfaces
curl http://localhost:11434/api/tags

Step 3: Deploy LiteLLM Proxy¶

LiteLLM provides an OpenAI-compatible API interface for Ollama models:

# Create LiteLLM configuration
cat > litellm_config.yaml <<EOF
model_list:
  - model_name: granite
    litellm_params:
      model: ollama/granite3.3
      api_base: http://host.docker.internal:11434
  - model_name: llama
    litellm_params:
      model: ollama/llama3.2
      api_base: http://host.docker.internal:11434
  - model_name: mistral
    litellm_params:
      model: ollama/mistral
      api_base: http://host.docker.internal:11434

general_settings:
  master_key: "sk-1234567890"  # Change this!
  database_url: "postgresql://openwebui:changeme@postgres:5432/litellm"
EOF

# Run LiteLLM container
docker run -d \
  --name litellm \
  --network openwebui-net \
  -v $(pwd)/litellm_config.yaml:/app/config.yaml \
  -p 4000:4000 \
  ghcr.io/berriai/litellm:main \
  --config /app/config.yaml \
  --detailed_debug

# Test LiteLLM
curl http://localhost:4000/v1/models \
  -H "Authorization: Bearer sk-1234567890"

Step 4: Set Up MCP Gateway¶

Deploy the MCP Gateway (ContextForge) for managing MCP servers:

# Create data directory
mkdir -p $(pwd)/mcpgateway_data

# Run MCP Gateway
docker run -d \
  --name mcpgateway \
  --network openwebui-net \
  -p 4444:4444 \
  -v $(pwd)/mcpgateway_data:/data \
  -e MCPGATEWAY_UI_ENABLED=true \
  -e MCPGATEWAY_ADMIN_API_ENABLED=true \
  -e DATABASE_URL=sqlite:////data/mcp.db \
  -e HOST=0.0.0.0 \
  -e JWT_SECRET_KEY=your-secret-key \
  -e BASIC_AUTH_USER=admin \
  -e BASIC_AUTH_PASSWORD=changeme \
  -e AUTH_REQUIRED=true \
  ghcr.io/ibm/mcp-context-forge:0.6.0

# Generate an API token for later use
docker exec mcpgateway \
  python3 -m mcpgateway.utils.create_jwt_token \
  --username admin --exp 0 --secret your-secret-key

Access the MCP Gateway UI at http://localhost:4444/admin (admin/changeme).

Step 5: Deploy MCPO¶

MCPO exposes MCP servers as OpenAPI-compatible endpoints:

# Install MCPO using UV (recommended)
pip install uv
uvx mcpo --version

# Or install with pip
pip install mcpo

# Create MCPO configuration
cat > mcpo_config.json <<EOF
{
  "mcpServers": {
    "time": {
      "command": "uvx",
      "args": ["mcp-server-time", "--local-timezone=America/New_York"]
    },
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/tmp"]
    },
    "memory": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-memory"]
    }
  }
}
EOF

# Run MCPO with configuration
uvx mcpo \
  --port 8000 \
  --api-key "mcpo-secret-key" \
  --config mcpo_config.json

# Test MCPO endpoints
curl http://localhost:8000/docs
curl http://localhost:8000/time/docs
curl -H "Authorization: Bearer mcpo-secret-key" \
  http://localhost:8000/time/list-tools

Step 6: Deploy OpenWebUI¶

Now deploy OpenWebUI with all integrations configured:

# Create OpenWebUI environment file
cat > openwebui.env <<EOF
# Database
DATABASE_URL=postgresql://openwebui:changeme@postgres:5432/openwebui

# OAuth (optional - adjust for your provider)
ENABLE_OAUTH_SIGNUP=false
ENABLE_SIGNUP=true
WEBUI_AUTH=true

# Model Backend (LiteLLM)
OPENAI_API_BASE_URL=http://litellm:4000/v1
OPENAI_API_KEY=sk-1234567890

# UI Settings
WEBUI_NAME=AI Assistant Platform
DEFAULT_MODELS=granite,llama,mistral

# MCP Tools URL (via MCPO)
TOOLS_FUNCTION_CALLING_PROMPT="Use available tools to assist with user requests"
EOF

# Run OpenWebUI
docker run -d \
  --name openwebui \
  --network openwebui-net \
  --env-file openwebui.env \
  -p 3000:8080 \
  -v openwebui_data:/app/backend/data \
  --restart always \
  ghcr.io/open-webui/open-webui:latest

# Check logs
docker logs -f openwebui

⚙️ Configuration¶

6.1 Configure OpenWebUI Model Connection¶

1. Access OpenWebUI at http://localhost:3000
2. Create an admin account on first login
3. Navigate to Settings → Connections
4. Verify the OpenAI connection is configured:
5. API Base URL: http://litellm:4000/v1
6. API Key: sk-1234567890

6.2 Register MCP Tools in OpenWebUI¶

1. Go to Settings → Tools
2. Click Add Tool Server
3. Add MCPO endpoints:

   Name: MCP Time Tool
   URL: http://host.docker.internal:8000/time
   API Key: mcpo-secret-key
   

4. Repeat for other MCPO tools (filesystem, memory, etc.)

6.3 Enable Tools for Models¶

1. Navigate to Workspace → Models
2. Edit each model (granite, llama, mistral)
3. Scroll to Tools section
4. Toggle on the MCP tools you want available
5. Click Save

6.4 Add MCP Servers to Gateway¶

1. Access MCP Gateway at http://localhost:4444/admin
2. Navigate to Gateways tab
3. Click Add Gateway to register external MCP servers
4. Example configuration:

   Name: Local MCP Time Server
   URL: http://host.docker.internal:8000/time
   Auth Type: Bearer Token
   Token: mcpo-secret-key
   

🧪 Testing the Integration¶

Test 1: Model Connectivity¶

# Test LiteLLM is serving models
curl -X POST http://localhost:4000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-1234567890" \
  -d '{
    "model": "granite",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

Test 2: MCP Tool Invocation¶

In OpenWebUI chat: 1. Enable the time tool by clicking the ➕ icon 2. Ask: "What's the current time in New York?" 3. The assistant should use the MCP time tool to respond

Test 3: Multi-Tool Workflow¶

Try complex queries that require multiple tools: - "Save a note about today's meeting at 3 PM EST" - "What files are in the /tmp directory?"

🔧 Advanced Configuration¶

Adding Custom MCP Servers¶

1. Install any MCP server:

   npm install -g @your-org/mcp-server-custom
   

2. Add to MCPO config:

   {
     "mcpServers": {
       "custom": {
         "command": "mcp-server-custom",
         "args": ["--option", "value"]
       }
     }
   }
   

3. Restart MCPO and register in OpenWebUI

Using Functions and Filters¶

OpenWebUI supports custom functions for extending capabilities:

1. Go to Settings → Functions
2. Enable built-in functions or import custom ones
3. Functions can:
4. Act as custom models (Pipe Functions)
5. Modify inputs/outputs (Filter Functions)
6. Add action buttons (Action Functions)

Performance Tuning¶

For production deployments:

# litellm_config.yaml additions
general_settings:
  max_parallel_requests: 100
  request_timeout: 600
  stream_timeout: 60

router_settings:
  routing_strategy: "usage-based-routing"
  redis_host: "redis"
  redis_port: 6379

📊 Monitoring and Maintenance¶

Container Health Checks¶

# Check all containers
docker ps --format "table {{.Names}}\t{{.Status}}\t{{.Ports}}"

# View logs
docker logs --tail 50 -f openwebui
docker logs --tail 50 -f litellm
docker logs --tail 50 -f mcpgateway

# Resource usage
docker stats

Backup and Restore¶

# Backup PostgreSQL
docker exec postgres pg_dump -U openwebui openwebui > backup.sql

# Backup volumes
docker run --rm -v postgres_data:/data -v $(pwd):/backup \
  alpine tar czf /backup/postgres_backup.tar.gz -C /data .

# Restore PostgreSQL
docker exec -i postgres psql -U openwebui openwebui < backup.sql

🚨 Troubleshooting¶

Common Issues¶

1. Models not appearing in OpenWebUI - Check Ollama is running: curl http://localhost:11434/api/tags - Verify LiteLLM can reach Ollama: docker logs litellm - Ensure model names match in litellm_config.yaml

2. MCP tools not working - Verify MCPO is running: curl http://localhost:8000/docs - Check tool registration in OpenWebUI settings - Ensure tools are enabled for the model

3. Container connectivity issues - Use host.docker.internal for host access from containers - Ensure all containers are on the same network - Check firewall rules

Debug Commands¶

# Test inter-container connectivity
docker exec openwebui ping postgres
docker exec openwebui curl http://litellm:4000/health

# Check MCPO tool availability
curl -H "Authorization: Bearer mcpo-secret-key" \
  http://localhost:8000/{tool-name}/list-tools

# Verify MCP Gateway
curl -H "Authorization: Bearer $(cat mcpgateway_token.txt)" \
  http://localhost:4444/tools

🎯 Next Steps¶

1. Add More Models: Explore Ollama's model library
2. Integrate More MCP Servers: Browse MCP servers directory
3. Customize UI: Modify OpenWebUI themes and settings
4. Set Up Authentication: Configure OAuth for production use
5. Enable Web Search: Add search capabilities to models
6. Document Indexing: Set up RAG with document collections

📚 Resources¶

• OpenWebUI Documentation
• Ollama Documentation
• LiteLLM Documentation
• MCPO Documentation
• MCP Gateway Documentation
• Model Context Protocol Spec

🐳 Docker Compose Deployment¶

For easier management, use Docker Compose:

# docker-compose.yml
version: '3.8'

services:
  postgres:
    image: postgres:15-alpine
    environment:
      POSTGRES_DB: openwebui
      POSTGRES_USER: openwebui
      POSTGRES_PASSWORD: changeme
    volumes:
      - postgres_data:/var/lib/postgresql/data
    networks:
      - openwebui-net
    healthcheck:
      test: ["CMD-SHELL", "pg_isready -U openwebui"]
      interval: 10s
      timeout: 5s
      retries: 5

  litellm:
    image: ghcr.io/berriai/litellm:main
    command: --config /app/config.yaml --detailed_debug
    volumes:
      - ./litellm_config.yaml:/app/config.yaml
    ports:
      - "4000:4000"
    networks:
      - openwebui-net
    depends_on:
      postgres:
        condition: service_healthy
    extra_hosts:
      - "host.docker.internal:host-gateway"

  mcpgateway:
    image: ghcr.io/ibm/mcp-context-forge:0.6.0
    environment:
      MCPGATEWAY_UI_ENABLED: "true"
      MCPGATEWAY_ADMIN_API_ENABLED: "true"
      DATABASE_URL: "sqlite:////data/mcp.db"
      HOST: "0.0.0.0"
      JWT_SECRET_KEY: "your-secret-key"
      BASIC_AUTH_USER: "admin"
      BASIC_AUTH_PASSWORD: "changeme"
      AUTH_REQUIRED: "true"
    volumes:
      - mcpgateway_data:/data
    ports:
      - "4444:4444"
    networks:
      - openwebui-net

  openwebui:
    image: ghcr.io/open-webui/open-webui:latest
    env_file:
      - openwebui.env
    ports:
      - "3000:8080"
    volumes:
      - openwebui_data:/app/backend/data
    networks:
      - openwebui-net
    depends_on:
      - postgres
      - litellm
    extra_hosts:
      - "host.docker.internal:host-gateway"

volumes:
  postgres_data:
  mcpgateway_data:
  openwebui_data:

networks:
  openwebui-net:
    driver: bridge

Deploy with:

docker-compose up -d
docker-compose logs -f

🔒 Security Best Practices¶

1. Use Strong Secrets¶

# Generate secure tokens
openssl rand -hex 32  # For JWT secrets
openssl rand -base64 32  # For API keys

2. Enable HTTPS¶

For production, use a reverse proxy with SSL:

server {
    listen 443 ssl http2;
    server_name your-domain.com;

    ssl_certificate /path/to/cert.pem;
    ssl_certificate_key /path/to/key.pem;

    location / {
        proxy_pass http://localhost:3000;
        proxy_set_header Host $host;
        proxy_set_header X-Real-IP $remote_addr;
        proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
        proxy_set_header X-Forwarded-Proto $scheme;
    }
}

3. Network Isolation¶

# docker-compose.yml addition
services:
  openwebui:
    networks:
      - frontend
      - backend

networks:
  frontend:
    driver: bridge
  backend:
    driver: bridge
    internal: true

4. Resource Limits¶

services:
  openwebui:
    deploy:
      resources:
        limits:
          cpus: '2'
          memory: 4G
        reservations:
          cpus: '1'
          memory: 2G