1. Click on "Install Server". 2. Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state. 3. In the chat, type @ followed by the MCP server name and your instructions, e.g., "@ns-hpc Create a new sandbox workspace for testing." That's it! The server will respond to your query, and you can continue using it as needed. Here is a step-by-step guide with screenshots.

ns-hpc — Namespaced HPC MCP Server

A Python-based Model Context Protocol (MCP) server that provides a secure, sandboxed interface for LLM Agents to interact with an HPC cluster. The core isolation mechanism is bubblewrap (bwrap) for unprivileged user namespace isolation.

Architecture

LLM Agent (Claude, etc.)
    │  MCP over STDIO (SSH)
    ▼
┌─────────────────────────────┐
│      ns-hpc MCP Server      │
│  ┌───────────────────────┐  │
│  │   Managed MCP Proxy   │──┼──► child MCP servers (filesystem, git, …)
│  └───────────────────────┘  │       inside bwrap container
│  ┌───────────────────────┐  │
│  │    Instance Manager   │──┼──► ~/mcp_instances/{id}/workspace/
│  └───────────────────────┘  │       + metadata.json + audit.log
│  ┌───────────────────────┐  │
│  │      Task Manager     │──┼──► local (Popen + bwrap)
│  │                       │  │    or Slurm (sbatch + bwrap)
│  └───────────────────────┘  │
└─────────────────────────────┘

Requirements

Python ≥ 3.11
bubblewrap (bwrap) — install: apt install bubblewrap or dnf install bubblewrap
User namespaces enabled — sysctl kernel.unprivileged_userns_clone=1
Slurm (optional) — for submitting jobs to the cluster

Quick Start

# Install ns-hpc
cd ns-hpc
uv sync

# Run diagnostics
uv run ns-hpc doctor

# Start the MCP server (over STDIO — connect via SSH)
uv run ns-hpc run

CLI

ns-hpc run       # Start the MCP server over STDIO
ns-hpc doctor    # Check bwrap, namespaces, and Slurm availability
ns-hpc --version # Show version

Configuration

Edit config.toml to customize:

namespace_defaults — bwrap flags (read-only dirs, dev/proc/tmpfs, environment)
proxied_mcps — child MCP servers to spawn inside bwrap containers
resource_defaults — Slurm walltime, CPUs, memory defaults
data_dir — where instance workspaces are stored (default: ~/mcp_instances)
context_dir — directory with Markdown documentation files

MCP Tools

Tool	Description
`create_instance`	Create a bwrap-sandboxed workspace
`destroy_instance`	Remove a workspace and all its data
`list_instances`	List all active instances
`read_audit_log`	Read the audit trail for an instance
`run_command`	Execute a command inside bwrap (local or Slurm)
`get_task`	Query task status and output
`list_tasks`	List tasks for an instance
`cancel_task`	Cancel a running task
`list_context_files`	List available HPC documentation
`read_context_file`	Read a documentation file

Security

All commands run inside a bwrap user namespace — no root required.
The audit log is written by the host process, never from inside the sandbox.
Network is disabled by default (--unshare-net).
Instance isolation via per-instance directories with workspace bind mounts.

Project Structure

ns-hpc/
├── config.toml              # Main configuration
├── context/                 # HPC documentation (exposed as MCP resources)
├── pyproject.toml           # Project metadata & dependencies
└── src/ns_hpc/
    ├── __init__.py
    ├── __main__.py
    ├── cli.py               # CLI entry point (run, doctor)
    ├── config.py            # Pydantic config models + TOML loader
    ├── server.py            # MCP server with all tool handlers + proxy
    └── core/
        ├── bwrap_builder.py  # bwrap argument list construction
        ├── instance_manager.py # Workspace CRUD + audit log
        └── task_manager.py  # Local & Slurm task execution

ns-hpc

ns-hpc — Namespaced HPC MCP Server

Architecture

Requirements

Quick Start

CLI

Configuration

MCP Tools

Security

Project Structure

Resources

Looking for Admin?

Latest Blog Posts

MCP directory API