MCP Protocol Proxy

Model Context Protocol (MCP) Proxy Interceptor 🔌

The Model Context Protocol (MCP) enables AI agents to call local and cloud tools via JSON-RPC. Without a security boundary, agents can pass sensitive environment variables or private file contents across MCP tool calls.

---

🔒 How AgentShield Protects MCP

AgentShield functions as an inline JSON-RPC proxy sitting between the agent client (Claude Desktop, Cursor, Antigravity) and the MCP server.

sequenceDiagram
    participant Agent as AI Coding Agent
    participant Shield as AgentShield MCP Proxy
    participant Server as MCP Tool Server

    Agent->>Shield: JSON-RPC call (tool: read_file, args: {path: ".env"})
    Shield->>Shield: Inspect tool arguments & validate AST
    Note over Shield: Detects restricted file target!
    Shield-->>Agent: JSON-RPC error ("Access to .env denied by AgentShield policy")

Loading

---

⚙️ Configuration in Claude Desktop / Cursor

Add the AgentShield proxy to your claude_desktop_config.json:

{
  "mcpServers": {
    "shielded-filesystem": {
      "command": "npx",
      "args": ["agentshield", "mcp-proxy", "--upstream", "@modelcontextprotocol/server-filesystem", "/workspace"]
    }
  }
}

Protection Guarantees:

1. Tool Argument Sanitization: In-flight strings containing secrets are substituted with surrogate tokens before reaching the tool.
2. Response Scrubbing: Output returning from MCP servers is audited for leaked cloud keys before entering the LLM prompt window.