VaikoraVaikora

Vaikora › Blog › Product Updates

Why policy enforcement belongs at the tool call, not after it: a walkthrough of Vaikora Guard MCP

Product Updates · Nov 14, 2026 · 4 min read

When an AI agent calls a tool, the action executes before most security controls have a chance to intervene. The gap is not theoretical. An agent reading a file, writing to a database, or posting to an external API does so in the same millisecond window that your logs are still buffering. By the time a detection rule fires, the data has already moved.

Vaikora Guard MCP closes that gap by sitting inline between the agent and every tool it calls. The enforcement decision, allow, block, modify, or require approval, happens before execution. Not after.

What the proxy actually does

The MCP (Model Context Protocol) is the emerging standard for how agents discover and call tools. Most MCP servers do no inspection at all. A tool call comes in, the server routes it to the underlying function, and the result comes back. There is no policy layer in that path.

Vaikora Guard MCP inserts itself into that path as a proxy. The agent sends its tool call to Vaikora Guard MCP rather than directly to the MCP server. The proxy evaluates the call against configured policy, then either passes it through, blocks it, rewrites it (the modify decision), or holds it for a human reviewer before forwarding.

The proxy is written in Python and published on PyPI as `vaikora-guard-mcp`. It is MIT licensed, self-hosted, and available on PyPI today. No vendor lock-in. You can read the full source, fork it, and deploy it inside your own infrastructure.

A concrete scenario

Say you have an agent configured to help engineers query internal documentation and also has access to a tool that can write to an S3 bucket. The agent is working correctly most of the time. But prompt injection from a retrieved document tells it to exfiltrate a file by writing it to a public bucket.

Without a policy gate, that write call executes. With Vaikora Guard MCP in the path, the tool call hits the proxy first. If your policy says "block any S3 write operation that targets a bucket outside the approved list," the call is blocked before it reaches S3. The agent receives a denial. The audit log records the decision, the input, and the policy that matched, all in a SHA-256 hash-chained, tamper-evident format.

That audit chain matters for compliance. Because each decision is deterministic (identical input always produces identical output) and signed into the chain, you can demonstrate to an auditor exactly what the agent attempted and exactly why it was blocked. Vaikora ships compliance presets for SOC 2, HIPAA, GDPR, PCI-DSS, and ISO 27001 so you are not writing those policy conditions from scratch.

What you configure

Policy in Vaikora Guard MCP is defined at the tool level. You specify:

Enforcement latency runs at P50 8ms and P99 45ms, with throughput above 10,000 actions per second. For most agent workflows, the proxy is not the bottleneck.

How it connects to the broader control plane

The open-source proxy is the enforcement point. The Vaikora Control Plane is the commercial layer that adds centralized policy management, the tamper-evident audit log at scale, compliance reporting, and integrations with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.

The split matters for how teams adopt it. A security engineer can deploy the open-source proxy today in front of an existing agent, start collecting decisions, and evaluate what policy gaps exist before committing to any commercial infrastructure. The Control Plane adds governance at the organizational level.

For teams already running in Azure or AWS, Vaikora is available as a SaaS subscription on AWS Marketplace with EDP draw-down support, and as the Vaikora-AzureSecurityCenter solution in the Azure Sentinel content hub.

What to check next

If you have agents running against MCP servers today, the first practical step is to map what tools those agents can call and whether any of those tools touch production data, external APIs, or privileged infrastructure. Then pull `vaikora-guard-mcp` from PyPI, run it in front of one agent in audit-only mode, and review what calls it logs before you write any blocking rules. The log will show you what your agents are actually doing.

See Vaikora enforce policy on your AI

Open-core AI runtime control. Self-host the MIT gateway free, or run the hosted Control Plane.

Get a demo Self-host the gateway

More from the Vaikora blog