Installation

Estimated time: ~10 minutes

What is PML?

PML (Procedural Memory Layer) is an intelligent MCP gateway that consolidates all your MCP servers into a single entry point with semantic search, DAG workflow orchestration, and self-improving tool discovery.

Problems solved:

  • Context saturation - Tool schemas consume 30-50% of LLM window → reduced to <5%
  • Sequential latency - Multi-tool workflows run serially → parallelized via DAG

Prerequisites

Before starting, make sure you have:

  • Deno 2.x or higher - Deno Installation
  • Git - To clone the repository
  • A coding agent - Claude Code, Cursor, or another MCP client

Verify Deno

deno --version

You should see:

deno 2.x.x (...)

Installation

Step 1: Clone the repository

git clone https://github.com/casys-ai/casys-pml.git
cd casys-pml

Step 2: Build the CLI

deno task build

You should see:

Compile file:///.../src/main.ts to pml

Step 3: Verify installation

./pml --help

Expected output:

Usage: pml [options] [command]

Commands:
  init    Initialize PML from MCP config
  serve   Start PML MCP gateway server
  status  Show gateway status and health

Tutorial: Your First Workflow with Claude Code

Let's configure PML as an MCP gateway for Claude Code in a few steps.

1. Prepare your MCP configuration

Create a configuration file for your MCP servers:

mkdir -p config
cat > config/mcp-servers.json << 'EOF'
{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-server-filesystem", "/path/to/allowed/dir"]
    },
    "memory": {
      "command": "npx",
      "args": ["-y", "@anthropic/mcp-server-memory"]
    }
  }
}
EOF

Tip: You can also migrate your existing Claude Desktop config with ./pml init --config ~/.config/Claude/claude_desktop_config.json

2. Initialize PML

./pml init --config config/mcp-servers.json

This command:

  • Discovers all your configured MCP servers
  • Extracts tool schemas via MCP protocol
  • Generates embeddings for semantic search
  • Stores everything in a local PGlite database (~/.pml/db)

Expected output:

🚀 Initializing PML...
✓ Found 2 MCP server(s)
✓ Extracted 15 tool schemas
✓ Generated embeddings (BGE-M3)
✓ Stored in ~/.pml/db

PML is ready!

3. Configure Claude Code

Add PML to your Claude Code MCP configuration:

Linux/macOS: ~/.config/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "pml": {
      "command": "/absolute/path/to/pml",
      "args": ["serve", "--config", "/absolute/path/to/config/mcp-servers.json"]
    }
  }
}

Important: Use absolute paths for command and --config.

4. Launch and test

Restart Claude Code. The gateway starts automatically.

To test manually:

stdio mode (default - recommended for Claude Code):

./pml serve --config config/mcp-servers.json

HTTP mode (with Fresh dashboard):

./pml serve --config config/mcp-servers.json --port 3001

Note: The Fresh dashboard (deno task dev:fresh) requires HTTP mode (--port). In stdio mode, only the MCP interface is available.

You should see:

🚀 Starting PML MCP Gateway...

Step 1/6: Loading configuration...
✓ Found MCP config: config/mcp-servers.json
Step 2/6: Initializing database...
Step 3/6: Connecting to MCP servers...
  ✓ Connected: filesystem
  ✓ Connected: memory
Step 4/6: Loading AI models...
Step 5/6: Starting MCP gateway...
Step 6/6: Listening for MCP requests...

PML gateway running on port 3001

Congratulations! You have configured PML as an intelligent MCP gateway.

First Steps with Meta-Tools

Once connected, test these tools in Claude Code:

Use pml:search_tools to find tools related to "read JSON files"

DAG workflow execution

Use pml:execute_dag with intent "Read config.json and create a memory entity"

Sandbox code execution

Use pml:execute_code to filter and aggregate data locally

Monitoring (optional)

PML includes a Grafana/Loki/Promtail stack for log monitoring:

# Start the monitoring stack
cd monitoring && docker-compose up -d

# Access Grafana (admin/admin)
open http://localhost:3000

Note: Monitoring works in both stdio AND Streamable HTTP modes because Promtail reads log files (~/.pml/logs/).

Next Steps

Now that you're up and running: