# Linux
curl -fsSL https://ollama.com/install.sh | sh

# macOS
brew install ollama

# Windows — download installer from https://ollama.com/download

# Verify installation
ollama --version

# Start the Ollama daemon
ollama serve

# It runs on http://localhost:11434 by default
# Keep this terminal open, or run as a system service:
# sudo systemctl enable ollama && sudo systemctl start ollama  (Linux)

# Recommended: good balance of speed and quality
ollama pull llama3.1:8b

# Alternatives:
ollama pull mistral              # Fast, 7B params
ollama pull codellama:13b        # Best for code tasks
ollama pull qwen2.5:7b           # Good multilingual
ollama pull llama3.1:70b         # Best quality (needs 40GB+ RAM)
ollama pull deepseek-coder:6.7b  # Specialized for code
ollama pull phi3:mini             # Smallest, fastest

# Verify model is available
ollama list

cd multi-agent-app

# Create Python virtual environment
python3 -m venv .venv

# Activate it
source .venv/bin/activate        # Linux/macOS
# .venv\Scripts\activate         # Windows PowerShell
# .venv\Scripts\activate.bat     # Windows CMD

# Install all dependencies
pip install -r requirements.txt

OLLAMA_BASE_URL = "http://localhost:11434"  # Ollama address
MODEL_NAME = "llama3.1:8b"                  # Model you pulled
MAX_TOKENS = 2048                           # Max response length

# Interactive CLI mode
python main.py

# Or as MCP server
python main.py --mcp-server

# List installed models
ollama list

# Pull a new model
ollama pull <model-name>

# Remove a model
ollama rm <model-name>

# Show model details
ollama show llama3.1:8b

# Test a model directly
ollama run llama3.1:8b "Hello, how are you?"

/model codellama:13b
/model mistral

# Change Ollama host/port (if needed)
export OLLAMA_HOST=0.0.0.0:11434

# Set GPU layers (for partial GPU offload)
export OLLAMA_NUM_GPU=999

# Set number of threads
export OLLAMA_NUM_THREAD=8

# stdio transport (for Claude Desktop, Cursor, etc.)
python main.py --mcp-server

# SSE/HTTP transport (for web clients or remote access on LAN)
python main.py --mcp-server --transport sse --host 0.0.0.0 --port 8080

python main.py --mcp-server [OPTIONS]

Options:
  --transport {stdio,sse}   Transport protocol (default: stdio)
  --host HOST               Bind address for SSE (default: 0.0.0.0)
  --port PORT               Port for SSE (default: 8080)

# Install Node.js MCP servers (one-time, cached locally)
npx -y @modelcontextprotocol/server-filesystem /tmp
npx -y @modelcontextprotocol/server-sqlite mydb.sqlite
npx -y @modelcontextprotocol/server-memory

# Or install Python-based MCP servers
pip install mcp-server-fetch
pip install mcp-server-git

EXTERNAL_MCP_SERVERS = [
    # Filesystem access — agents can read/write local files
    {
        "name": "filesystem",
        "transport": "stdio",
        "command": "npx",
        "args": ["-y", "@modelcontextprotocol/server-filesystem", "/home/jaspal/projects"]
    },

    # SQLite — agents can query local databases
    {
        "name": "sqlite",
        "transport": "stdio",
        "command": "npx",
        "args": ["-y", "@modelcontextprotocol/server-sqlite", "/home/jaspal/data/app.db"]
    },

    # Memory/Knowledge base — persistent agent memory
    {
        "name": "memory",
        "transport": "stdio",
        "command": "npx",
        "args": ["-y", "@modelcontextprotocol/server-memory"]
    },

    # Git — agents can interact with local repos
    {
        "name": "git",
        "transport": "stdio",
        "command": "python",
        "args": ["-m", "mcp_server_git", "--repo", "/home/jaspal/projects/myapp"]
    },

    # Custom local MCP server (running on localhost)
    {
        "name": "custom-tools",
        "transport": "sse",
        "url": "http://localhost:9090/sse"
    },
]

┌─────────────────────────────────────────────────────────┐
│                    YOUR MACHINE                          │
│                                                         │
│  ┌─────────────┐     ┌──────────────────────────────┐  │
│  │   Ollama    │     │   Multi-Agent App            │  │
│  │ (Local LLM) │◄───►│   45 agents + supervisor     │  │
│  │ :11434      │     │   MCP server (stdio/SSE)     │  │
│  └─────────────┘     └──────────┬───────────────────┘  │
│                                  │                      │
│                    ┌─────────────┼─────────────┐        │
│                    ▼             ▼             ▼        │
│  ┌──────────────┐ ┌───────────┐ ┌───────────────┐     │
│  │ MCP Server:  │ │MCP Server:│ │ MCP Server:   │     │
│  │ filesystem   │ │ sqlite    │ │ memory        │     │
│  │ (local files)│ │ (local db)│ │ (local store) │     │
│  └──────────────┘ └───────────┘ └───────────────┘     │
│                                                         │
│  ┌──────────────────────────────────────────────────┐  │
│  │ MCP Clients: Claude Desktop / Cursor / VS Code   │  │
│  └──────────────────────────────────────────────────┘  │
│                                                         │
│  Network: ZERO external traffic                         │
└─────────────────────────────────────────────────────────┘

# Terminal 1: Start Ollama
ollama serve

# Terminal 2: Run the app
cd multi-agent-app
source .venv/bin/activate
python main.py

🧑 You: write a Python REST API with Flask for a todo app

/workflow full_dev implement user authentication with JWT and refresh tokens

/workflow bug_fix TypeError: Cannot read property 'map' of undefined in UserList.tsx

/workflow design design a real-time notification system for 100k users

/workflow docs document the payment processing module with API reference

/files src/auth.py,src/middleware.py security audit these files

/workflow perf_audit our API response time is 2s, analyze and optimize

/ask coder write a Python decorator for caching with TTL
/ask database design a schema for multi-tenant SaaS
/ask devops write a GitHub Actions CI/CD pipeline for a Node.js app
/ask shell_expert write a bash script to backup PostgreSQL daily

/feedback coder reviewer write a thread-safe LRU cache in Python

/compare architect,optimizer design a caching strategy for product catalog

/parallel security,optimizer,accessibility audit this React component

/batch coder implement stack;;implement queue;;implement linked list;;implement BST

/auto our login endpoint is returning 500 errors in production

/files src/api.py,src/models.py,src/tests.py review for consistency issues

/workflow learn explain event-driven architecture with examples
/ask mentor explain the CAP theorem like I'm a junior developer

/workflow release prepare v2.0 release for the payment service

/workflow startup_mvp build a SaaS invoicing app with Stripe integration

{
  "mcpServers": {
    "multi-agent": {
      "command": "/home/jaspal/jscode/js-ai-apps-api/multi-agent-app/.venv/bin/python",
      "args": ["/home/jaspal/jscode/js-ai-apps-api/multi-agent-app/main.py", "--mcp-server"]
    }
  }
}

{
  "multi-agent": {
    "command": "/home/jaspal/jscode/js-ai-apps-api/multi-agent-app/.venv/bin/python",
    "args": ["main.py", "--mcp-server"],
    "cwd": "/home/jaspal/jscode/js-ai-apps-api/multi-agent-app"
  }
}

{
  "mcp": {
    "servers": {
      "multi-agent": {
        "command": "python",
        "args": ["main.py", "--mcp-server"],
        "cwd": "/home/jaspal/jscode/js-ai-apps-api/multi-agent-app"
      }
    }
  }
}

# Start HTTP server
python main.py --mcp-server --transport sse --port 8080

# Connect from any MCP client to: http://localhost:8080/sse

/feedback coder reviewer write a production-ready connection pool

/remember project Using PostgreSQL 15 with pgvector extension
/remember style snake_case, type hints, 4-space indent
/recall project
/forget project

/parallel security,optimizer,reviewer analyze this module

/batch tester write tests for login;;signup;;logout;;password-reset

/auto deploy our app to kubernetes with monitoring

/files src/api.py,src/models.py,tests/test_api.py find inconsistencies

/chain planner|architect|coder|tester|documentation build a rate limiter

/save my-project
/load my-project
/export

/tokens
# Output: Tokens: ~12,450 total (4,200 in / 8,250 out) | Requests: 7

export OLLAMA_HOST=0.0.0.0:11434   # Bind address
export OLLAMA_NUM_GPU=999           # GPU layers
export OLLAMA_NUM_THREAD=8          # CPU threads
export OLLAMA_KEEP_ALIVE=5m         # Model keep-alive time

multi-agent-app/
├── agent_registry.py   # 45 agent definitions (single source of truth)
├── config.py           # All settings + supervisor prompt
├── graph.py            # LangGraph supervisor orchestration
├── runners.py          # Execution modes + 50 workflows + advanced features
├── session.py          # History, tokens, save/load/export
├── main.py             # CLI dispatcher + entry point
├── mcp_server.py       # FastMCP server (dynamic tool registration)
├── tool_registry.py    # External MCP server consumption
├── requirements.txt    # Python dependencies
└── sessions/           # Saved sessions + agent memory
    ├── *.json          # Session files
    ├── *.md            # Exported conversations
    └── memory.json     # Persistent agent memory

"my_agent": {
    "description": "What it does",
    "temperature": 0.3,
    "prompt": "You are a ... agent. Your job is to: 1) ... 2) ... 3) ...",
},

"my_workflow": ["planner", "my_agent", "reviewer", "tester"],

{"name": "my-server", "transport": "stdio", "command": "python", "args": ["my_server.py"]}

# Check if running
curl http://localhost:11434/api/tags

# Start it
ollama serve

# List available models
ollama list

# Pull the model
ollama pull llama3.1:8b

# Make sure venv is activated
source .venv/bin/activate

# Reinstall dependencies
pip install -r requirements.txt

langchain>=0.3.0
langchain-ollama>=0.2.0
langgraph>=0.2.0
pydantic>=2.0.0
mcp>=1.0.0
requests>=2.28.0