Model Context Protocol (MCP) Server

A secure, multi‑tenant server that provides context to LLMs using an MCP-inspired protocol. Features JWT authentication, row‑level security (RLS), and audit logging.

---

Table of Contents

• Overview
• Architecture
• Features
• Tech Stack
• Prerequisites
• Setup & Installation
• Running the Server
• API Endpoints
• Testing with LLM
• Security Considerations
• Audit Logs
• Troubleshooting
• Next Steps

---

Overview

This project demonstrates a secure context server that can be integrated with LLMs (like Ollama) to provide real‑time, tenant‑specific data. It mimics the Model Context Protocol (MCP) concept, where an LLM requests context from backend systems in a safe, auditable way.

Key aspects:

• Authentication: JWT tokens identify the tenant.
• Authorization: PostgreSQL RLS ensures tenants only see their own data.
• Audit: Every context request is logged.
• Simplicity: The server is API‑only; the LLM can call it via a tool.

---

Architecture

graph TD
    subgraph Client
        A[LLM / Agent] --> B[FastAPI Server]
    end

    subgraph MCP Server
        B --> C[JWT Auth<br/>Extract Tenant]
        C --> D[PostgreSQL<br/>with RLS]
        D --> E[Context Data]
        B --> F[Audit Log<br/>PostgreSQL]
    end

    subgraph External
        G[Ollama LLM] --> B
    end

    E --> B
    B --> A

To generate a PNG image, copy the Mermaid code into mermaid.live and export as PNG.

---

Features

• JWT Authentication: Tokens contain tenant_id claim.
• Multi‑Tenant Data: Each tenant sees only their own orders and users.
• Row‑Level Security: PostgreSQL RLS enforces tenant isolation.
• Audit Logging: All context requests are logged with timestamp, tenant, and endpoint.
• Sample Data: Pre‑loaded synthetic customers and orders for tenants tenant_a and tenant_b.
• LLM Integration Example: Script shows how an LLM (via Ollama) can call the context server.

---

Tech Stack

Component	Technology
Server	Python + FastAPI
Database	PostgreSQL with RLS
Authentication	JWT (PyJWT)
Audit	Custom PostgreSQL table
Container	Docker Compose

---

Prerequisites

• Python 3.10+
• Docker and Docker Compose
• Ollama (optional, for testing LLM integration)

---

Setup & Installation

1. Clone the Repository

git clone https://github.com/your-username/mcp-context-server.git
cd mcp-context-server

2. Create Virtual Environment

python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

3. Install Dependencies

pip install -r requirements.txt

4. Start PostgreSQL

docker-compose up -d

5. Configure Environment

Copy .env.example to .env and edit if needed (defaults are fine for local).

6. Initialize Database

python scripts/init_db.py

This creates tables, enables RLS, inserts sample tenants and data.

7. Generate Test Tokens

python scripts/generate_token.py --tenant tenant_a

Copy the token output. You'll use it in API requests.

---

Running the Server

Start the FastAPI server:

uvicorn src.context_service:app --reload --port 8000

The API will be available at http://localhost:8000.

---

API Endpoints

GET /health

Health check.

GET /context/orders

Returns orders for the authenticated tenant.

Headers:

Authorization: Bearer <JWT>

Response (example):

[
  {"id": 1, "customer_name": "Alice", "total": 1200.0},
  {"id": 2, "customer_name": "Bob", "total": 850.0}
]

GET /context/customers

Returns customers for the authenticated tenant.

Audit Logs

All requests are logged in the audit_logs table. You can inspect them:

psql -h localhost -U postgres -d mcp_db -c "SELECT * FROM audit_logs ORDER BY timestamp DESC LIMIT 5;"

---

Testing with LLM

You can test how an LLM (like Ollama) can call this context server. Example script:

# example_llm_call.py
import requests
import json

JWT = "your_generated_token"
API_URL = "http://localhost:8000/context/orders"

response = requests.get(
    API_URL,
    headers={"Authorization": f"Bearer {JWT}"}
)
orders = response.json()
print("Orders:", orders)

# Now feed this context into an LLM (e.g., via Ollama)
context = f"Orders: {json.dumps(orders)}"
# Call Ollama with a prompt using the context...

You can extend this to a full agent that decides which endpoint to call based on the user's question.

---

Security Considerations

• JWT secret: Store securely, use a strong key.
• PostgreSQL RLS: Ensures even if a tenant obtains another tenant's JWT (unlikely with proper signing), they can't access other data.
• Audit: Logs all requests for compliance.
• TLS: In production, use HTTPS.

---

Audit Logs

All context requests are logged in the audit_logs table with:

• timestamp
• tenant_id
• endpoint
• user_id (optional, can be extended)

To view recent logs:

psql -h localhost -U postgres -d mcp_db -c "SELECT * FROM audit_logs ORDER BY timestamp DESC LIMIT 10;"

---

Troubleshooting

Problem	Solution
Token invalid	Check the JWT secret in .env matches the one used in generation.
No data returned	Verify the tenant ID in the token exists in the tenants table.
RLS errors	Ensure you enabled RLS on tables and created policies correctly (the init script does this).
PostgreSQL connection refused	Check docker-compose ps; ensure container is running.

---

Next Steps

• Add more context endpoints (e.g., GET /context/user/{id}).
• Integrate with LangChain as a custom tool.
• Deploy to cloud with managed PostgreSQL.
• Add rate limiting and request throttling.

---

License

MIT