mcpai-agentstool-usellm-security
MCP Server Action Gate: Verify Tools Before Execution
Gate every MCP tool call — approve, verify, or block before the server executes filesystem, API, and device side effects.
May 21, 20267 min read
Model Context Protocol (MCP) gives LLMs structured access to tools — files, databases, browsers, smart devices. Each tool call is an execution event. Sanctum gates MCP actions before your server runs them.
Why MCP needs a runtime gate
- Tools can delete files, exfiltrate secrets, or call production APIs
- Indirect prompt injection arrives through tool output, not user chat
- Multiple clients may share one MCP server — policies must be per actor and org
Integration pattern
// Before MCP tool executes:
const result = await sanctum.verifyAction({
actor: 'mcp-host',
action: 'write_file',
context: { path: '/etc/hosts', source_trust: 'tool_output' },
})
if (result.decision !== 'APPROVED') return heldOrBlockedSanctum ships an MCP marketplace adapter. See docs and the GitHub repo.
Related: source-trust classification, action approval.
Guides: agentic AI risk · MCP security · runtime authorization · HITL approvals · coding agents · get started
More: all posts · AI trust layer · open Sanctum Console