agent-wall wrap

Wrap an MCP server with Agent Wall policy enforcement.

Usage

agent-wall wrap [options] -- <command> [args...]

Options

Option	Description
-c, --config <path>	Path to agent-wall.yaml config file
-l, --log-file <path>	Path to write audit log (JSON lines)
-s, --silent	Suppress Agent Wall output (only MCP protocol on stdout)
--dry-run	Preview policy evaluation without starting the server
-d, --dashboard	Launch the real-time security dashboard
--dashboard-port <port>	Dashboard port (default: 61100)

Also respects AGENT_WALL_CONFIG and AGENT_WALL_LOG environment variables as fallbacks.

Examples

# Basic usage
agent-wall wrap -- npx @modelcontextprotocol/server-filesystem /home/user

# Custom config
agent-wall wrap -c custom-policy.yaml -- node my-server.js

# With audit logging  
agent-wall wrap --log-file ./audit.log -- npx mcp-server

# Silent mode (for Claude Desktop or non-interactive contexts)
agent-wall wrap --silent -- npx mcp-server

# Dry-run: preview rules without starting
agent-wall wrap --dry-run -- npx mcp-server

# Launch with real-time dashboard
agent-wall wrap --dashboard -- npx mcp-server

# Dashboard on custom port
agent-wall wrap --dashboard --dashboard-port 8080 -- node my-server.js

How it works

1. Loads the policy configuration (from file or built-in defaults)
2. Creates the response scanner from the responseScanning config section
3. Initializes security modules (injection detection, egress control, kill switch, chain detection)
4. Spawns the MCP server as a child process
5. Starts the dashboard server if --dashboard is specified
6. Intercepts all stdin/stdout JSON-RPC messages
7. Inbound: Runs the 5-step defense pipeline (kill switch → injection → egress → policy → chain)
8. Forwards allowed calls, blocks denied calls, prompts for approval
9. Outbound: Scans tool call responses for secrets, PII, and oversize content
10. Blocks or redacts responses with findings, passes clean responses through
11. Passes all other messages through untouched

Dry Run Mode

The --dry-run flag shows your loaded policy without starting the server:

╔══════════════════════════════════════════════════╗
║  Agent Wall — Security Firewall for AI Agents    ║
╠══════════════════════════════════════════════════╣
║  Policy: ./agent-wall.yaml
║  Rules:  8 loaded
║  Scanner: response scanning ON
║  Server: npx mcp-server
╚══════════════════════════════════════════════════╝

  --dry-run mode: preview only, server will not start

  Default action: prompt
  Rate limit:     200 calls / 60s
  Rules loaded:   8

  ✗ block-ssh-keys → deny (tool: *)
  ✗ block-exfiltration → deny (tool: shell_exec|run_command)
  ? approve-shell → prompt (tool: shell_exec|bash)
  ✓ allow-reads → allow (tool: read_file|get_file_contents)