Claude Agent SDK Safety Hooks

Claude's agent capabilities work through the Model Context Protocol (MCP), where tools are exposed as MCP servers. Authensor's MCP gateway sits between Claude and your MCP servers, enforcing policies on every tool call. This guide covers the integration patterns.

MCP Gateway Setup

Authensor's MCP server acts as a safety-aware proxy. Configure Claude to connect to Authensor's MCP server, which in turn connects to your actual tool servers. Every tool call passes through policy evaluation.

{
  "mcpServers": {
    "authensor": {
      "command": "npx",
      "args": ["@authensor/mcp-server", "--config", "./authensor.yaml"]
    }
  }
}

The gateway exposes the same tools as the downstream servers but wraps each with policy enforcement.

Policy Configuration for Claude Agents

Write policies that govern which tools Claude can access and under what conditions. Use Authensor's YAML policy format to define rules per tool, per agent, or globally.

Common rules for Claude agents include: limiting file system access to specific directories, restricting network requests to approved domains, requiring human approval for destructive operations, and blocking shell command execution in production.

Tool Description Security

MCP tool descriptions are part of the model's context. Authensor's gateway sanitizes tool descriptions to prevent injection attacks through tool metadata. It also validates that tool descriptions match a registered schema, catching attempts to modify tool behavior through description manipulation.

Human-in-the-Loop Integration

For sensitive operations, configure Authensor to pause execution and request human approval. The MCP gateway returns a pending status to Claude, which can inform the user that approval is needed. Once approved, the tool call proceeds.

This integrates with Claude's built-in confirmation patterns. The agent explains what it wants to do, the human approves through Authensor's approval interface, and execution continues.

Audit Trail

Every tool call Claude makes is recorded as an immutable receipt. The audit trail captures the tool name, arguments, policy decision, execution result, and timing. This gives you complete visibility into what Claude did and why.

Multi-Agent Patterns

When Claude delegates to sub-agents or other MCP servers, each hop through the gateway gets its own policy evaluation and audit record. The receipt chain links related actions across agents, giving you end-to-end traceability.