Castling Systems
Model Context Protocol Governance

Agent Governance for MCP Systems

Policy enforcement, smart approvals, and audit trails for AI agents

Public Beta / PoC Stage
AI
Agent
Castling
Tools

Agent → Policy Gateway → MCP Tools

The Agentic Governance Gap

Traditional network controls lack visibility into agent-level operations

The "Black Box" Problem

Standard WAFs cannot parse MCP JSON-RPC. You don't know if an agent is deleting a file or just reading it.

Indirect Prompt Injection

A malicious email or PDF can hijack your agent and force it to execute unauthorized tools.

Permissions Sprawl

Agents often have "Superuser" access. Once inside, they can do anything the API allows.

How It Works: ACE

Policy enforcement, tool masking, and human oversight in a single gateway

A

ANALYZE (Context & Session)

Evaluate session context and risk signals before execution. Policy rules run on every tool call.

C

CONSTRAIN (Dynamic Masking)

Agents see only tools they need for the current task. Limit exposure, reduce attack surface.

E

ESCALATE (Human-in-the-Loop)

Route high-risk actions to your existing approval workflow. Don't block valid work — approve it intelligently.

Built for Restricted Environments

Deep packet inspection, audit evidence, and on-prem deployment

What It Does

MCP Deep Inspection

Inspect MCP JSON-RPC requests and tool-call payloads in-line. Enforce allow/deny/step-up policies and capture evidence.

Policy-as-Code

Versioned, reviewable policies (Git-friendly) for tool access, environments, and approval rules.

Audit Evidence

Structured audit trails for incident response and compliance reviews (exportable JSON).

How It Runs

On-Prem / Air-Gapped

Designed for restricted networks with no required outbound connectivity.

Packaging

Delivered as a Dockerized gateway. Built on .NET for performance and native .NET ecosystem compatibility.

Current deployment model: Central MCP Gateway (inline). Distributed enforcement is roadmap.

Policy-as-Code Examples

Allow read-only filesystem access, block writes to protected paths

{
  "name": "restrict_filesystem_writes",
  "effect": "Deny",
  "resources": {
    "tools": ["filesystem.*"]
  },
  "conditions": "resource.toolName.startsWith('write') || resource.path.contains('/etc/') || resource.path.contains('/config/')"
}

What You Get

Clear value for every stakeholder in your organization

For CISO

  • Reduce agent risk with policy enforcement
  • Complete audit trail for compliance

For CTO

  • Ship agents with built-in guardrails
  • Policy-as-code for version control

For DevOps

  • Single gateway for all agent traffic
  • Smart escalation reduces approval fatigue

For Engineers

  • Ship agents faster with built-in safety
  • Clear feedback on blocked actions

Example Scenarios

Rogue MCP Server

Allowlist trusted servers, quarantine unknown ones. Block unauthorized tool invocations in real-time.

Prompt Injection

Malicious input hijacks agent. Policy rules block unauthorized tool sequences and escalate for review.

High-Impact Tools

Delete, deploy, payment operations require human approval with full audit trail and step-up verification.

Limited Availability

Shaping the Standard for Agent Governance

We are accepting 3 Enterprise Partners for Q2 Pilots

What is it?

A guided PoC where we deploy Castling to secure your specific agentic workflow. Work directly with our engineering team to build policies tailored to your use case.

Who is this for?

Teams running MCP agents in production (or near-prod) who need auditability, compliance, or are handling sensitive data with AI agents.

Included

  • Deployment support & integration guidance
  • Custom policy authoring for your workflow
  • Direct R&D channel with the founding team

Request Pilot Access

Tell us about your agentic workflow