Agent Workspace MCP

🛡️ Agent Workspace MCP Server

A unified Model Context Protocol (MCP) server providing a highly secure, containerized workspace for Large Language Models (LLMs). It acts as an isolated "agentic playground" where agents can autonomously code, test, and debug without risking the host machine.

✨ Features

• 🏗️ Full Project Lifecycle: Bootstrap projects with uv init, manage dependencies with uv add, and execute via uv run.
• 🐚 Secure Bash Access: Execute shell commands with mandatory timeouts and merged output streams.
• 🚀 Token-Optimized Output: Integrates RTK (Rust Token Killer) to automatically filter and compress run_bash outputs (like ls, git, and test runners), saving 60-90% of LLM context tokens.
• 📂 Robust Filesystem: Path-traversal protected operations for reading, writing, and searching the workspace.
• 🛡️ Multi-Layer Security: Non-root execution, dropped capabilities, resource limits, and a read-only root filesystem.
• ⚡ Precision Editing: Advanced search_and_replace with fuzzy whitespace matching, indentation preservation, dry-run support, and syntax validation for Python, JSON, JSONL, TOML, and YAML.
• 📊 Real-time Observability: Direct logging to MCP client UI and persistent rotating audit logs.

🏗️ Architecture

flowchart TD
    Client["MCP Client (Claude / Cursor)"] -- "stdio (JSON-RPC)" --> FastMCP["FastMCP Server"]

    subgraph Sandbox ["Docker Sandbox Container (mcpuser)"]
        direction TB
        
        FastMCP -. "Intercepts accidental prints" .-> StdioGuard["StdoutRedirector"]
        FastMCP -. "Application Logs" .-> Logger["Dual Logger (stderr & .mcp/server.log)"]
        
        FastMCP -- "Tool Calls" --> SecurityGuard["Security & Path Validator"]
        
        subgraph Toolset ["Tool Modules"]
            direction TB
            SecurityGuard --> FSTools["Filesystem (read, write, list, search)"]
            SecurityGuard --> EditTools["Editing (search_and_replace)"]
            SecurityGuard --> ExecTools["Execution (run_bash)"]
        end

        EditTools -- "AST Verification" --> Validator["Syntax Validations (Python, JSON, JSONL, TOML, YAML)"]
        ExecTools -- "Process Group (Timeout=60s)" --> Shell["/bin/sh Subprocess"]
        Shell -- "Package Mgt & Checks" --> UV["uv Environment / Ruff"]
        
        FSTools -- "Secure I/O" --> Workspace["/workspace Directory"]
        EditTools -- "Atomic Writes" --> Workspace
        Shell -- "Executes within" --> Workspace
    end

    Workspace <--"Volume Mount"--> HostFS["User Host Filesystem"]

📦 Quick Start

1. Pull or Build the Docker Image

# Pull from GHCR
docker pull ghcr.io/hrrodan/agent-workspace-mcp:latest

# OR: Build locally with your host's UID/GID for optimal permissions
docker build --build-arg UID=$(id -u) --build-arg GID=$(id -g) -t agent-workspace-mcp .

2. Programmatic Usage (OpenAI Agents SDK)

Here is a quick boilerplate showing how to use the containerized workspace programmatically using the standard openai-agents SDK:

import asyncio
from agents import Agent, Runner
from agents.mcp import MCPServerStdio

async def main():
    # 1. Configure the MCP Server to run via Docker
    server = MCPServerStdio(
        name="Sandboxed Workspace",
        params={
            "command": "docker",
            "args": [
                "run", "-i", "--rm", "--init",
                # "--network", "none", # Network Isolation (optional) - see below
                "--memory=2g", "--cpus=2.0",
                "--pids-limit=256",
                "--cap-drop=ALL", "--security-opt=no-new-privileges:true",
                "--read-only",
                "--tmpfs", "/tmp:size=64m",
                "--tmpfs", "/home/mcpuser/.cache:size=512m",
                "--user", "1000:1000", # Replace with your host UID:GID
                "-v", "/path/to/your/projects:/workspace",
                "ghcr.io/hrrodan/agent-workspace-mcp:latest",
            ],
        },
        client_session_timeout_seconds=60.0,
    )

    # 2. Attach server to the Agent and load the skill instructions (optional)
    with open("skills/agent-workspace-mcp/SKILL.md", "r") as f:
        skill_instructions = f.read()

    agent = Agent(
        name="WorkspaceAgent",
        instructions=f"You are a coding agent with access to a secure workspace.\n\n{skill_instructions}",
        mcp_servers=[server],
    )

    # 3. Execute a workflow
    async with server:
        result = await Runner.run(
            agent, 
            "Create a python script in the workspace to print the first 10 Fibonacci numbers, then run it."
        )
        print(f"Agent's Final Output:\n{result.final_output}")

if __name__ == "__main__":
    asyncio.run(main())

3. Use with MCP Clients (Claude / Cursor)

Add the following configuration to your claude_desktop_config.json or Cursor settings.

{
  "mcpServers": {
    "agent-workspace-mcp": {
      "command": "docker",
      "args": [
        "run", "-i", "--rm", "--init",
        // "--network", "none", // Network Isolation (optional) - see below
        "--memory=2g", "--cpus=2.0",
        "--pids-limit=256",
        "--cap-drop=ALL", "--security-opt=no-new-privileges:true",
        "--read-only",
        "--tmpfs", "/tmp:size=64m",
        "--tmpfs", "/home/mcpuser/.cache:size=512m",
        "--user", "1000:1000",
        "-v", "/path/to/your/projects:/workspace",
        "ghcr.io/hrrodan/agent-workspace-mcp:latest"
      ]
    }
  }
}

[!IMPORTANT] Linux Users: Replace 1000:1000 with your actual UID:GID (run id -u and id -g). Claude Desktop does not expand environment variables. Signal Handling: The --init flag is essential for proper signal forwarding and zombie process reaping.

🛠️ Tool Reference

⚙️ Configuration

The server supports the following environment variables (passed via Docker --env):

🛡️ Security & Architecture Model

This server employs a defense-in-depth strategy, explicitly separating strict security boundaries from developer experience and operational reliability features.

🔒 Core Security Features

These features are designed to protect the host system and enforce strict isolation boundaries.

• Kernel Hardening: All Linux capabilities are dropped (--cap-drop=ALL), neutralizing privilege escalation vectors.
• Immutable Server Code: The /app directory containing the server source and its virtual environment is owned by root and read-only for the mcpuser. This prevents the server from modifying itself or being tampered with via run_bash.
• Privilege Lockdown: Enforces no-new-privileges:true to prevent any process from gaining elevated rights.
• Immutable System Core: The container's root filesystem is mounted entirely read-only, providing a second layer of defense against OS-level tampering.
• Resource Quotas: Hard limitations on CPU, Memory, and PIDs mitigate denial-of-service (DoS) attempts like fork-bombs and host exhaustion.
• Strict Boundary Enforcement: A robust path validator comprehensively blocks all path traversal attacks outside the designated /workspace.
• Process & Resource Control: Mandatory command timeouts (default 60s) and strict process group isolation ensure runaway or malicious processes are killed.
• Memory-Overload Protection: Hard limits on file reads (1MB) and command outputs (50KB) prevent memory exhaustion.
• Information Leakage Prevention: Internal stack traces and system paths are suppressed and sanitized from tool outputs.

🛠️ Developer Experience & Convenience

Features focused on seamless integration, usability, and reducing friction during agentic workflows.

• Host-Aligned Non-Root Identity: Runs as mcpuser with UID/GID customizable at build time, eliminating tedious file permission conflicts on host volume mounts.
• Automatic Token Optimization: Shell commands executed via run_bash are transparently rewritten through RTK to provide ultra-compact, LLM-friendly output without altering underlying command behavior.
• Intelligent Search Exclusions: High-noise or sensitive directories (.git, .venv) are automatically ignored to keep context windows lean and relevant.
• Ephemeral Workspaces: Containers are strictly ephemeral (--rm), guaranteeing a clean, predictable slate for every new session without state leaking across connections.
• Standardized Discovery: Complies with the OCI Image Specification for standardized container ecosystem integration and transparent auditing.

⚙️ Reliability & Safety Mechanisms

Features ensuring the structural integrity of the workspace and providing observability.

• Pre-Write Syntax Validation: Both write_file and search_and_replace perform in-memory syntax validation for Python, JSON, JSONL, TOML, and YAML before persisting changes, preventing broken code states.
• Fail-Safe Writing: write_file blocks accidental overwrites of existing files by default and enforces a 5MB size guard to prevent workspace flooding.
• Atomic File Operations: Edits utilize temp-and-move logic to guarantee file integrity and prevent corruption, even during unexpected interruptions or crashes.
• Transparent Observability: All tool invocations and state changes are streamed in real-time to the MCP client UI for immediate operator oversight.

🌐 Network Isolation (Optional)

By default, the container has full network access via Docker's bridge network. For maximum isolation, you can completely disable the network stack using --network none:

docker run -i --rm --init \
  --network none \
  --memory=2g --cpus=2.0 --pids-limit=256 \
  --cap-drop=ALL --security-opt=no-new-privileges:true \
  --read-only \
  --tmpfs /tmp:size=64m \
  --tmpfs /home/mcpuser/.cache:size=512m \
  --user 1000:1000 \
  -v /path/to/your/projects:/workspace \
  ghcr.io/hrrodan/agent-workspace-mcp:latest

This creates a fully air-gapped sandbox — only the loopback interface exists inside the container. All outbound connections (curl, DNS, uv add, etc.) will fail immediately, eliminating data exfiltration and lateral movement risks entirely.

[!NOTE] With --network none, the agent cannot install packages at runtime. All dependencies must be pre-installed in a custom image or pre-populated in the mounted workspace volume.

🤝 Contributing

1. Install Dev Dependencies: uv sync
2. Run Linting: uv run ruff check .
3. Run Unit Tests: uv run pytest tests/ --ignore=tests/integration/
4. Run Integration Tests: Set OPENROUTER_API_KEY and run uv run pytest tests/integration/

© 2026 HrRodan. Licensed under MIT.