Step 1: Choose your integration method

Install one or both:

# Option A: MCP proxy (intercepts all tool calls)
npm install -g tyga-mcp-proxy

# Option B: CLI wrapper (prefix commands manually)
curl -fsSL https://a2ainfrastructure.com/install.sh | bash

Step 2: Add the safety layer

Option A: MCP Proxy

Add to your editor's MCP configuration file:

// Cursor:   ~/.cursor/mcp.json
// Windsurf: ~/.windsurf/mcp.json
// Devin:    configured in workspace settings
{
  "mcpServers": {
    "tyga-firewall": {
      "command": "tyga-mcp-proxy",
      "args": ["--upstream", "npx -y @anthropic-ai/mcp-shell"],
      "env": {
        "A2A_API_KEY": "a2a_your_key_here"
      }
    }
  }
}

Every tools/call is intercepted and screened. Dangerous commands never reach the shell.

Option B: CLI Wrapper

Configure your editor to use tyga as the shell prefix:

# In your terminal, prefix commands with tyga
tyga npm install
tyga python train.py
tyga git push origin main

# Or set as default shell wrapper in editor settings:
# "terminal.integrated.shell.args": ["tyga"]

Step 3: Verify

# In your editor's terminal or AI chat:

# Ask the AI to list files -> ALLOWED
"Show me all TypeScript files in src/"
# -> executes normally

# Ask the AI to delete files -> BLOCKED
"Remove node_modules and all build artifacts recursively"
# -> "Command blocked by safety shield: destructive operation"