MCP DevBench

Production-ready Docker container management server implementing the Model Context Protocol (MCP)

MCP DevBench provides isolated, persistent development workspaces through a secure, audited, and observable container management API. Built for AI assistants like Claude, it enables safe command execution and filesystem operations in Docker containers.

---

✨ Features

Core Capabilities

• 🚀 Container Lifecycle Management - Create, start, stop, and remove Docker containers with fine-grained control
• 📁 Secure Filesystem Operations - Read, write, delete files with path validation and ETag-based concurrency control
• ⚡ Async Command Execution - Non-blocking execution with streaming output and timeout handling
• 🔐 Enterprise Security - Capability dropping, read-only rootfs, resource limits, and comprehensive audit logging
• 📊 Production Observability - Prometheus metrics, structured JSON logging, and system health monitoring
• ⚙️ True Async I/O - All blocking operations wrapped in thread pools for optimal concurrency

Advanced Features

• Warm Container Pool - Sub-second container provisioning for instant attach
• Graceful Shutdown - Drain active operations before server termination
• Automatic Recovery - Reconciles Docker state with database on startup
• Image Policy Enforcement - Allow-list validation with digest pinning
• Multi-Transport Support - stdio, SSE, or HTTP-based MCP transports
• Flexible Authentication - None, Bearer token, or OIDC authentication modes

---

🚀 Quick Start

Prerequisites

• Python 3.11+
• Docker Engine
• uv package manager

Installation

# Install uv
pip install uv

# Clone the repository
git clone https://github.com/pvliesdonk/mcp-devbench.git
cd mcp-devbench

# Install dependencies
uv sync

Running the Server

Development Mode (stdio)

uv run python -m mcp_devbench.server

Production Mode (HTTP)

export MCP_TRANSPORT_MODE=streamable-http
export MCP_HOST=0.0.0.0
export MCP_PORT=8000
uv run python -m mcp_devbench.server

Using Docker

docker build -t mcp-devbench .
docker run -v /var/run/docker.sock:/var/run/docker.sock \
  -p 8000:8000 \
  -e MCP_TRANSPORT_MODE=streamable-http \
  mcp-devbench

Using Docker Compose

docker-compose up -d

---

🏗️ Architecture

┌─────────────────────────────────────────────────────┐
│                   MCP DevBench API                   │
│           (FastMCP Server with Auth)                │
└─────────────────┬───────────────────────────────────┘
                  │
    ┌─────────────┼─────────────┐
    │             │             │
    ▼             ▼             ▼
┌─────────┐  ┌─────────┐  ┌──────────┐
│Container│  │  Exec   │  │Filesystem│
│ Manager │  │ Manager │  │ Manager  │
└────┬────┘  └────┬────┘  └────┬─────┘
     │            │            │
     └────────────┼────────────┘
                  │
         ┌────────┼────────┐
         │        │        │
         ▼        ▼        ▼
    ┌────────┐ ┌────┐ ┌──────────┐
    │ Docker │ │ DB │ │  Audit   │
    │ Daemon │ │    │ │  Logger  │
    └────────┘ └────┘ └──────────┘

Design Patterns:

• Repository Pattern - Data access abstraction
• Manager Pattern - Business logic encapsulation
• Dependency Injection - Loose coupling via factory functions
• Async/Await - Non-blocking I/O throughout with thread pool for blocking operations

---

📚 Usage Examples

Basic Container Workflow

from mcp_devbench.mcp_tools import *

# 1. Spawn a container
result = await spawn(SpawnInput(
    image="python:3.11-slim",
    persistent=False,
    alias="dev-workspace"
))
container_id = result.container_id

# 2. Attach to container
await attach(AttachInput(
    target=container_id,
    client_name="my-client",
    session_id="session-123"
))

# 3. Execute command
exec_result = await exec_start(ExecInput(
    container_id=container_id,
    cmd=["python", "--version"],
    timeout_s=30
))

# 4. Poll for output
output = await exec_poll(ExecPollInput(
    exec_id=exec_result.exec_id,
    after_seq=0
))

# 5. Write a file
await fs_write(FileWriteInput(
    container_id=container_id,
    path="/workspace/hello.py",
    content=b"print('Hello, World!')"
))

# 6. Clean up
await kill(KillInput(
    container_id=container_id,
    force=True
))

---

⚙️ Configuration

All configuration is managed through environment variables with the MCP_ prefix.

Essential Settings

| Variable | Default | Description | |----------|---------|-------------| | MCP_TRANSPORT_MODE | streamable-http | Transport: stdio, sse, or streamable-http | | MCP_HOST | 0.0.0.0 | Server bind address (HTTP transports only) | | MCP_PORT | 8000 | Server port (HTTP transports only) | | MCP_ALLOWED_REGISTRIES | docker.io,ghcr.io | Comma-separated allowed registries | | MCP_LOG_LEVEL | INFO | Logging level: DEBUG, INFO, WARNING, ERROR | | MCP_LOG_FORMAT | json | Log format: json or text |

Authentication

| Variable | Default | Description | |----------|---------|-------------| | MCP_AUTH_MODE | none | Auth mode: none, bearer, or oidc | | MCP_BEARER_TOKEN | - | Bearer token (when auth_mode=bearer) | | MCP_OAUTH_CLIENT_ID | - | OIDC client ID | | MCP_OAUTH_CLIENT_SECRET | - | OIDC client secret | | MCP_OAUTH_CONFIG_URL | - | OIDC discovery URL |

Advanced Settings

| Variable | Default | Description | |----------|---------|-------------| | MCP_STATE_DB | ./state.db | SQLite database path | | MCP_DRAIN_GRACE_S | 60 | Shutdown grace period (seconds) | | MCP_TRANSIENT_GC_DAYS | 7 | Transient container retention (days) | | MCP_WARM_POOL_ENABLED | true | Enable warm container pool | | MCP_DEFAULT_IMAGE_ALIAS | python:3.11-slim | Default warm pool image |

Example Configurations

Local Development (stdio)

MCP_TRANSPORT_MODE=stdio
MCP_AUTH_MODE=none
MCP_LOG_LEVEL=DEBUG
MCP_LOG_FORMAT=text

Production (HTTP + OIDC)

MCP_TRANSPORT_MODE=streamable-http
MCP_HOST=0.0.0.0
MCP_PORT=8000
MCP_AUTH_MODE=oidc
MCP_OAUTH_CLIENT_ID=your-client-id
MCP_OAUTH_CLIENT_SECRET=your-secret
MCP_OAUTH_CONFIG_URL=https://auth.example.com/.well-known/openid-configuration
MCP_LOG_LEVEL=INFO
MCP_LOG_FORMAT=json

---

🔧 MCP Tools Reference

Container Management

spawn

Create and start a new container.

Input:

• image (string) - Docker image reference
• persistent (boolean) - Persist across restarts
• alias (string, optional) - User-friendly name
• ttl_s (integer, optional) - Time-to-live for transient containers

Output:

• container_id (string) - Opaque container ID
• alias (string) - Container alias
• status (string) - Container status

attach

Attach a client to a container for session tracking.

Input:

• target (string) - Container ID or alias
• client_name (string) - Client identifier
• session_id (string) - Session identifier

Output:

• container_id (string) - Actual container ID
• alias (string) - Container alias
• roots (array) - Workspace roots

kill

Stop and remove a container.

Input:

• container_id (string) - Container to remove
• force (boolean) - Force immediate removal

Output:

• status (string) - Operation status

Command Execution

exec_start

Start command execution in a container.

Input:

• container_id (string) - Target container
• cmd (array) - Command and arguments
• cwd (string) - Working directory (default: /workspace)
• env (object) - Environment variables
• as_root (boolean) - Execute as root
• timeout_s (integer) - Execution timeout
• idempotency_key (string) - Prevent duplicate execution

Output:

• exec_id (string) - Execution ID
• status (string) - Initial status

exec_cancel

Cancel a running execution.

exec_poll

Poll for execution output and status.

Input:

• exec_id (string) - Execution ID
• after_seq (integer) - Return messages after sequence number

Output:

• messages (array) - Stream messages
• complete (boolean) - Execution complete flag

Filesystem Operations

fs_read

Read a file from container workspace.

Output includes: content, etag, size, mime_type

fs_write

Write a file to container workspace.

Supports: ETag-based concurrency control via if_match_etag

fs_delete

Delete a file or directory.

fs_stat

Get file/directory metadata.

fs_list

List directory contents.

System & Monitoring

system_status

Get system health and status.

Output:

• Docker connectivity status
• Active containers/attachments count
• Database status
• Server version

metrics

Retrieve Prometheus-formatted metrics.

reconcile

Manually trigger container reconciliation.

garbage_collect

Trigger manual garbage collection.

list_containers / list_execs

List all containers or active executions.

---

🔐 Security

Built-in Security Features

• Capability Dropping - All Linux capabilities dropped by default
• Read-Only Root Filesystem - Prevents container modification
• Resource Limits - 512MB memory, 1 CPU, 256 PID limit per container
• Path Validation - Prevents directory traversal attacks
• Image Allow-List - Only approved registries allowed
• Audit Logging - Complete audit trail with PII redaction
• User Isolation - Configurable UID (default 1000)

Security Best Practices

1. Use OIDC Authentication in production
2. Restrict allowed registries to trusted sources only
3. Enable audit logging and monitor for suspicious activity
4. Run with least privilege - never run as root
5. Keep images updated - use digest pinning for reproducibility
6. Isolate network access - use Docker network policies

---

📊 Observability

Structured Logging

All operations are logged in JSON format with:

• ISO8601 timestamps
• Correlation IDs
• Contextual metadata
• Automatic PII redaction

Prometheus Metrics

Available via the metrics tool:

• mcp_devbench_container_spawns_total - Container creation count
• mcp_devbench_exec_total - Command execution count
• mcp_devbench_exec_duration_seconds - Execution duration histogram
• mcp_devbench_fs_operations_total - Filesystem operation count
• mcp_devbench_active_containers - Active container gauge
• mcp_devbench_memory_usage_bytes - Container memory usage

Audit Events

All operations generate audit events:

• CONTAINER_SPAWN, CONTAINER_ATTACH, CONTAINER_KILL
• EXEC_START, EXEC_CANCEL
• FS_READ, FS_WRITE, FS_DELETE
• SYSTEM_RECONCILE, SYSTEM_GC

---

🧪 Development

For detailed development guidelines, see our Project Style Guide.

Running Tests

# Run all tests
uv run pytest

# Run with coverage
uv run pytest --cov=mcp_devbench --cov-report=html

# Run specific test file
uv run pytest tests/unit/test_container_manager.py

# Run integration tests only
uv run pytest tests/integration/

Code Quality

# Lint with ruff
uv run ruff check .

# Format code
uv run ruff format .

# Type checking (recommended)
uv run pyright src/

Project Structure

mcp-devbench/
├── src/mcp_devbench/
│   ├── config/          # Configuration management
│   ├── models/          # SQLAlchemy ORM models
│   ├── managers/        # Business logic layer
│   ├── repositories/    # Data access layer
│   ├── utils/           # Utilities (logging, Docker, metrics)
│   ├── server.py        # FastMCP server
│   └── mcp_tools.py     # Pydantic models for MCP
├── tests/
│   ├── unit/            # Unit tests
│   └── integration/     # Integration tests
├── alembic/             # Database migrations
└── .github/workflows/   # CI/CD pipelines

---

📈 Project Status

Current Version: 0.1.0

Completed Epics

✅ Epic 1: Foundation Layer - Configuration, state store, Docker lifecycle ✅ Epic 2: Command Execution - Async exec, streaming, idempotency ✅ Epic 3: Filesystem Operations - CRUD, batch ops, import/export ✅ Epic 4: MCP Integration - Tools, resources, streaming transport ✅ Epic 5: Security - Image policy, hardening, warm pool ✅ Epic 6: State Management - Shutdown, recovery, maintenance ✅ Epic 7: Observability - Audit logging, metrics, admin tools

Test Coverage: ~72% (201 tests) Code Quality: Zero linting issues (ruff) Production Ready: Yes, for small-to-medium deployments

Roadmap

See IMPLEMENTATION_ROADMAP.md for detailed future plans.

---

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details on:

• Development workflow
• Testing guidelines
• Code style requirements
• Submission process

Important: We use a main/next branching model. See BRANCHING_STRATEGY.md for details.

Quick Contribution Steps

1. Fork the repository
2. Create a feature branch from next: git checkout next && git pull && git checkout -b feature/amazing-feature
3. Make your changes and add tests
4. Run tests: uv run pytest
5. Lint code: uv run ruff check .
6. Commit with conventional commits: git commit -m "feat: add amazing feature"
7. Push and create a Pull Request to the next branch

---

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

---

🔗 Resources

• Documentation: docs/
• Issue Tracker: GitHub Issues
• Discussions: GitHub Discussions
• Changelog: CHANGELOG.md

---

💬 Support

• Questions? Open a Discussion
• Bug Reports: File an Issue
• Security Issues: See SECURITY.md

---

Built with ❤️ using FastMCP, Docker, and modern Python async