Skip to content

achimstruve/MCPAuthPrototype

BranchesTags

Repository files navigation

MCP Auth Prototype

A secure Model Context Protocol (MCP) server demonstrating enterprise-grade access control patterns. Built with Python and FastMCP v2, this prototype implements JWT-based authentication and scope-based tool authorization, designed for deployment on Google Kubernetes Engine.

What This Demonstrates

  • Token-based authentication: Every MCP request requires a valid JWT Bearer token
  • Scope-based authorization: Token scopes determine which tools a client can see and call
  • Defense in depth: Tool list filtering AND tool call validation (two independent checks)
  • Structured audit logging: Every auth decision logged as JSON for cloud logging systems
  • 12-factor configuration: All settings via environment variables
  • Kubernetes-ready: Health and readiness probe endpoints

Architecture

Client (Claude Code, MCP client)
  │
  │  Authorization: Bearer <jwt>
  ▼
┌─────────────────────────────────┐
│  FastMCP Server (port 8080)     │
│                                 │
│  ┌───────────────────────────┐  │
│  │  AuthMiddleware           │  │
│  │  1. Extract Bearer token  │  │
│  │  2. Validate JWT (sig+exp)│  │
│  │  3. Filter tools by scope │  │
│  │  4. Block unauthorized    │  │
│  └───────────────────────────┘  │
│                                 │
│  ┌───────────┐ ┌─────────────┐  │
│  │ get_public│ │get_confiden-│  │
│  │ _info     │ │tial_info    │  │
│  │           │ │             │  │
│  │ scope:    │ │ scope:      │  │
│  │ public:   │ │ confidenti- │  │
│  │ read      │ │ al:read     │  │
│  └───────────┘ └─────────────┘  │
│                                 │
│  /health  /ready  /mcp          │
└─────────────────────────────────┘

Access Control Matrix

Token Scopes Visible Tools Can Call
["public:read"] get_public_info only get_public_info only
["public:read", "confidential:read"] Both tools Both tools
[] None None
No token / expired / invalid Rejected (AuthError) Rejected (AuthError)

Quick Start

Prerequisites

  • Python 3.11+
  • uv package manager

Install and Run

# Install dependencies
uv sync

# Start the server
uv run python -m src.server

The server starts on http://localhost:8080 with:

  • MCP endpoint: POST /mcp (Streamable HTTP transport)
  • Health check: GET /health
  • Readiness check: GET /ready

Generate a Token

# Public access only
uv run python -m scripts.generate_token --sub alice --scope public:read

# Full access
uv run python -m scripts.generate_token --sub bob --scope public:read confidential:read

# Expired token (for testing rejection)
uv run python -m scripts.generate_token --sub charlie --scope public:read --exp-hours -1

Note: The token must be signed with the same secret the server uses. By default, both use dev-secret-change-me. If you run the server with a custom secret (e.g., MCP_JWT_SECRET_KEY=my-secret), you must generate tokens with the matching --secret flag:

uv run python -m scripts.generate_token --sub alice --scope public:read --secret my-secret

Connect with Claude Code

# Generate a token
TOKEN=$(uv run python -m scripts.generate_token --sub myuser --scope public:read confidential:read 2>&1 | grep "^Token:" | cut -d' ' -f2)

# Add the MCP server to Claude Code
claude mcp add --transport http mcp-auth-prototype http://localhost:8080/mcp \
  --header "Authorization: Bearer $TOKEN"

Test with curl

# Generate a token
TOKEN=$(uv run python -m scripts.generate_token --sub alice --scope public:read 2>&1 | grep "^Token:" | cut -d' ' -f2)

# Initialize MCP session
curl -X POST http://localhost:8080/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

Development

# Run tests
uv run pytest -v

# Lint
uv run ruff check .

Docker

Build and test the container image locally before deploying to Kubernetes.

Build the Image

# Build the Docker image
docker build -t mcp-auth-prototype:local .

The multi-stage build creates a minimal ~150MB image containing only the runtime dependencies.

Run the Container

# Run with a custom JWT secret (required for production)
docker run -p 8080:8080 -e MCP_JWT_SECRET_KEY=my-secret mcp-auth-prototype:local

# Run with debug logging
docker run -p 8080:8080 \
  -e MCP_JWT_SECRET_KEY=my-secret \
  -e MCP_LOG_LEVEL=debug \
  mcp-auth-prototype:local

Test the Container

# Verify health endpoint
curl http://localhost:8080/health

# Verify readiness endpoint
curl http://localhost:8080/ready

# Generate a token (must use --secret matching the container's MCP_JWT_SECRET_KEY)
TOKEN=$(uv run python -m scripts.generate_token --sub alice --scope public:read --secret my-secret 2>&1 | grep "^Token:" | cut -d' ' -f2)

# Test MCP initialization against the container
curl -X POST http://localhost:8080/mcp \
  -H "Content-Type: application/json" \
  -H "Accept: application/json, text/event-stream" \
  -H "Authorization: Bearer $TOKEN" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize","params":{"protocolVersion":"2025-03-26","capabilities":{},"clientInfo":{"name":"test","version":"1.0"}}}'

Project Structure

mcp-auth-prototype/
├── src/
│   ├── server.py          # MCP server, auth middleware, health endpoints
│   ├── auth.py            # JWT validation and scope extraction
│   ├── tools.py           # Tool-to-scope mapping registry
│   └── config.py          # Environment-based configuration (pydantic-settings)
├── documents/
│   ├── public.md          # Sample public company document
│   └── confidential.md    # Sample confidential strategy document
├── scripts/
│   └── generate_token.py  # CLI utility to mint JWT tokens
├── tests/
│   ├── conftest.py        # Shared test fixtures (token factories)
│   ├── test_auth.py       # Unit tests for JWT validation (16 tests)
│   └── test_tools.py      # Integration tests for tool authorization (6 tests)
├── helm/
│   └── mcp-server/            # Helm chart for Kubernetes deployment
│       ├── Chart.yaml         # Chart metadata (name, version)
│       ├── values.yaml        # Default configuration values
│       ├── values-dev.yaml    # Dev environment overrides
│       └── templates/
│           ├── _helpers.tpl       # Reusable Go template helpers
│           ├── deployment.yaml    # Deployment (2 replicas, probes, env vars)
│           ├── service.yaml       # ClusterIP Service on port 8080
│           ├── configmap.yaml     # Document content (public.md, confidential.md)
│           ├── serviceaccount.yaml # K8s ServiceAccounts with Workload Identity
│           ├── secretstore.yaml   # ESO connection to GCP Secret Manager
│           └── externalsecret.yaml # Syncs JWT key from GCP to K8s Secret
├── terraform/             # Infrastructure as Code
│   ├── main.tf            # Provider and backend configuration
│   ├── variables.tf       # Input variables
│   ├── outputs.tf         # Output values
│   ├── gke.tf             # GKE cluster definition
│   ├── artifact-registry.tf  # Container registry
│   ├── secret-manager.tf  # Secret Manager resources
│   ├── iam.tf             # Service accounts and IAM bindings
│   └── github-wif.tf     # Workload Identity Federation for GitHub Actions
├── .github/
│   └── workflows/
│       └── ci.yaml        # CI pipeline (lint, test, build, push, update Helm)
├── argocd/
│   └── application.yaml   # ArgoCD Application (GitOps auto-sync)
├── pyproject.toml         # Dependencies and tool configuration
└── uv.lock                # Locked dependency versions

CI/CD Pipeline

Every push to main triggers an automated pipeline:

git push ──▶ GitHub Actions ──▶ ArgoCD ──▶ GKE Cluster
              │                   │
              ├─ Lint (ruff)      ├─ Detects values.yaml change
              ├─ Test (pytest)    ├─ Renders Helm chart
              ├─ Build image      └─ Rolling update (zero downtime)
              ├─ Push to Artifact Registry (git SHA tag)
              └─ Update helm/mcp-server/values.yaml
  • No stored credentials: GitHub Actions authenticates to GCP via Workload Identity Federation (OIDC token exchange)
  • Immutable image tags: Docker images are tagged with the git commit SHA (e.g., a1b2c3d), not latest
  • GitOps: ArgoCD continuously reconciles the cluster state with what's in Git — including self-healing if someone manually modifies the cluster

Design Decisions

Document Storage: ConfigMap (Prototype) vs Production Alternatives

In this prototype, document content (public.md, confidential.md) is inlined directly in a Kubernetes ConfigMap within the Helm chart. This is appropriate here because:

  • We have only 2 small, static documents (~1KB total)
  • It keeps the Helm chart self-contained and easy to understand
  • Helm's .Files.Get function can't read files outside the chart directory

This approach does NOT scale. ConfigMaps are limited to 1MB, document changes require a full Helm upgrade (which triggers a pod rolling update), and there's no versioning or independent lifecycle management.

Production alternatives for document-heavy systems:

Pattern When to Use How It Works
Object Storage (GCS/S3) Most common. Independent document lifecycle, many documents App fetches from a cloud bucket at runtime via Workload Identity. Supports versioning, CDN, fine-grained IAM.
Database (PostgreSQL/Firestore) Documents need metadata, search, relationships App queries a database per request. Full CRUD, indexing, transactions.
Git repo + sidecar GitOps-heavy orgs, docs-as-code A sidecar/init container clones a separate docs repo. Version history from Git, PRs for review.
Content API microservice Large-scale, many consumers Dedicated service manages documents. MCP server becomes a thin orchestration layer.

The key principle: decouple document lifecycle from application lifecycle. The MCP server should be deployable independently from content updates.

Configuration

All settings are read from environment variables with the MCP_ prefix:

Variable Default Description
MCP_HOST 0.0.0.0 Network interface to bind to
MCP_PORT 8080 Server port
MCP_LOG_LEVEL info Logging verbosity (debug, info, warning, error)
MCP_JWT_SECRET_KEY dev-secret-change-me JWT signing key (override in production)
MCP_JWT_ALGORITHM HS256 JWT signing algorithm
MCP_DOCUMENTS_DIR documents Path to document files

You can also set these in a .env file (gitignored).

Technology Stack

Component Technology Purpose
MCP Server FastMCP v2 MCP protocol with middleware hooks
Authentication PyJWT JWT token validation
Configuration pydantic-settings Typed env var config
HTTP Server Uvicorn ASGI server
Testing pytest + httpx Unit and integration tests
Linting Ruff Fast Python linter
Package Manager uv Fast Python package manager
Infrastructure Terraform Infrastructure as Code for GCP resources
Container Registry GCP Artifact Registry Docker image storage
Orchestration Google Kubernetes Engine Container orchestration
Secrets GCP Secret Manager + ESO Secure secret management
CI GitHub Actions Automated lint, test, build, push
CD ArgoCD GitOps continuous deployment
CI→GCP Auth Workload Identity Federation OIDC-based auth, no stored keys

Roadmap

See IMPLEMENTATION_ROADMAP.md for the full build plan. Current status:

  • Phase 0: Project scaffolding
  • Phase 1: MCP server with tools
  • Phase 2: Authentication and authorization
  • Phase 3: Tests
  • Phase 4: Dockerize
  • Phase 5: GCP Infrastructure + Terraform + GKE
  • Phase 6: Helm chart
  • Phase 7: GitHub Actions CI pipeline
  • Phase 8: ArgoCD
  • Phase 9: End-to-end verification
  • Phase 10: TLS Ingress (HTTPS) — Ingress controller, cert-manager, Let's Encrypt, encrypted external access
  • Phase 11: OAuth2 Token Service — Production token issuance via Google OAuth2, developer CLI, Claude Code integration
  • Phase 12: Autoscaling & Resilience — HPA, Cluster Autoscaler, PDB, load balancing, load testing with Locust

About

A scalable MCP server prototype deployed in a Kubernetes cluster via GKE leveraging the GCP secret manager and a tiered authentication regime for tool access.

Resources

Readme
Activity

Stars

1 star

Watchers

0 watching

Forks

0 forks
Report repository

Releases

No releases published

Packages

 
 
 

Contributors

Languages