MCP Integration

Overview

AgentTrust provides mcp_tool_guard and MCPServerGuard for Model Context Protocol servers and tool invocations. Validates tool inputs/outputs at the MCP boundary.

Tier: Developer+

Why It Matters

MCP tools expose agents to external capabilities. Tool trust checks at the MCP layer prevent unauthorized or risky tool executions.

Prerequisites

pip install agentrust-py
# Developer+ API key for MCP adapter

Step-by-Step Guide

from mcp.server.fastmcp import FastMCP
from agentrust_sdk import mcp_tool_guard, MCPServerGuard

# ── Option A: guard individual tool functions ──────────────────────────
@mcp_tool_guard(agent_id="mcp-tools", base_url="http://127.0.0.1:8765")
async def my_tool_handler(request):
    return await execute_tool(request)


# ── Option B: guard every tool registered on a server ──────────────────
mcp = FastMCP("payments-server")

@mcp.tool()
async def process_payment(amount: float, account: str) -> dict:
    ...

# The server instance is the FIRST positional argument.
guard = MCPServerGuard(mcp, agent_id="payments-mcp", base_url="http://127.0.0.1:8765")
# process_payment is now governed — every call is validated.

MCPServerGuard patches the already-registered tool handlers, so construct it after all @mcp.tool() registrations. Both entry points default base_url to http://localhost:8000 and do not read AGENTRUST_GATEWAY_URL.

Examples

See agentrust_sdk/agentrust_sdk/adapters/mcp.py.

Best Practices

• Guard at server level for comprehensive coverage
• Include tool name and arguments in validation metadata
• Use tool trust policy rules for allowlist/denylist

Common Mistakes

• Catching TierGateError — the guard raises plain RuntimeError at construction
• Constructing MCPServerGuard(agent_id=...) without the server — server is the first positional argument
• Constructing the guard before the tools are registered

Troubleshooting

Related Resources

• Tool Trust (How It Works)
• Policy Packs