vlp-mcp-agent

MCP server for VLP lab VM automation. Runs on the manager VM of an EPC HOL SKU vPOD alongside the existing vlp-tools-agent. Exposes VM capabilities as first-class MCP tools so Cursor, Claude Code, and other AI agents can automate lab tasks without manual SSH sessions.

What it does

Wraps the co-located vlp-tools-agent REST API at localhost:8787 and adds general VM tooling on top. The existing agent's WebSocket connection to the VLP Hub is untouched.

Developer Machine
  └── SSH tunnel :8789  →  Manager VM :5480
                               ├── vlp-mcp-agent (this project)  :8789
                               │     └── calls localhost:8787
                               └── vlp-tools-agent (Spring Boot)  :8787
                                     └── outbound WSS → VLP Hub

Tools available to the AI agent

VLP-specific

General VM tools

Remote side — install once (manager VM)

git clone git@benhtodd-bc:ATE-Labs/vlp-mcp-agent.git
cd vlp-mcp-agent
./install.sh

Installs to /opt/vlp-mcp-agent/. Service is installed but not started.

Local side — connect when needed

Start / stop

# Connect (starts service on manager VM + opens SSH tunnel)
./connect-manager.sh holuser@<manager-ip>

# Disconnect (closes tunnel + stops service)
# Press Ctrl-C in the connect-manager.sh terminal — it handles cleanup automatically.
# Or manually:
./disconnect-manager.sh holuser@<manager-ip>

Environment variable overrides:

VLP_SSH_PORT=5480   # SSH port for the manager VM (default: 5480)
VLP_MCP_PORT=8789   # MCP port to tunnel (default: 8789)

On-demand via SSH (no tunnel, no persistent service)

For one-off use, Cursor can launch the server as a subprocess over SSH — no tunnel needed:

{
  "mcpServers": {
    "vlp-manager": {
      "command": "ssh",
      "args": ["-p", "5480", "holuser@<manager-ip>",
               "/opt/vlp-mcp-agent/venv/bin/python",
               "/opt/vlp-mcp-agent/vlp_mcp_agent/server.py",
               "--stdio"]
    }
  }
}

Agent config — connect via SSH tunnel

After connect-manager.sh is running:

Cursor

Edit ~/.cursor/mcp.json (global) or .cursor/mcp.json (project):

{
  "mcpServers": {
    "vlp-manager": {
      "url": "http://localhost:8789/sse"
    }
  }
}

Claude Code

Edit ~/.claude/settings.json:

{
  "mcpServers": {
    "vlp-manager": {
      "url": "http://localhost:8789/sse"
    }
  }
}

Development / local testing

# Install deps
python3 -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"   # or: pip install -r requirements.txt

# Run locally (no vlp-tools-agent required — tools that need it will report gracefully)
python vlp_mcp_agent/server.py --port 8789

# Run tests
pytest

Relationship to vlp-tools-agent

This project is a companion, not a replacement. The vlp-tools-agent continues to:

• Maintain the WebSocket connection to the VLP Hub

• Handle executeVmScript messages from VLP

• Manage reconnection, heartbeats, and pending message queues

This agent adds the MCP interface layer on top, enabling AI-driven automation. When the vlp-tools-agent owners add a POST /api/script/execute endpoint, execute_script will automatically use it; until then it falls back to direct subprocess execution.

Security

• Binds to 127.0.0.1 only — not reachable without SSH access to the manager VM

• SSH (port 5480) is the authentication layer — no separate token needed

• Configurable command timeouts and blocked path prefixes in config.yaml

• Restart=no in systemd unit — never auto-starts, must be started manually