VaikoraVaikora

Vaikora › Blog › Product Updates

What goes into vaikora-guard-mcp's policy check, and what the audit record looks like when it blocks

Product Updates · Nov 22, 2026 · 4 min read

Tool calls are the moment of actual risk in an MCP-based agent workflow. The model has decided what to do. The question is whether anything checks that decision before the underlying function executes. vaikora-guard-mcp is an open-source proxy that sits in that gap.

This article walks through what the proxy inspects, how a policy decision is recorded, and what you would see in the audit trail after a block.

Where the proxy sits in the call path

When an AI agent calls a tool through MCP, the message travels from the model to the MCP server, which routes it to the actual implementation. vaikora-guard-mcp inserts itself as a proxy between the model and the MCP server. Every tool call passes through it before the server sees the request.

The proxy is written in Python, published on PyPI as `vaikora-guard-mcp`, and licensed under MIT. You can read the source, modify the policy engine, and run it on your own infrastructure. Nothing phones home by default.

What the policy check evaluates

For each incoming tool call, the proxy evaluates four things in order:

The result is a deterministic decision: allow, block, modify, or require-approval. The same input always produces the same decision. There is no probabilistic scoring that might go a different way on the next run.

A concrete example: a rule that blocks any `write_file` call where the parameter `path` matches `/etc/*` will block every such call, every time, before the MCP server processes it. The file is never touched. The agent sees a policy rejection, not a permission error from the OS.

What the audit record contains

Every decision, including allowed calls, writes to a SHA-256 hash-chained audit log. The chain means you can verify that no entry was modified or deleted after the fact. Each record contains:

If a call is blocked and you later need to reconstruct what the agent attempted, the audit log gives you the exact parameters the model supplied, not a summary. That matters during an investigation. You are looking at what actually happened, not an approximation.

Writing a rule and seeing it fire

Policy rules in vaikora-guard-mcp are declarative. A rule that blocks database deletion calls outside business hours looks roughly like this:

```
- tool: delete_records
condition: hour < 8 or hour > 18
action: block
reason: "Deletion outside approved window"
```

When the proxy evaluates a matching call, the response back to the agent includes the reason string. The audit log entry records the rule name and the full condition that matched.

You do not need to modify the agent, change the MCP server, or update the underlying tool implementation. The rule lives in the proxy configuration. The proxy enforces it inline.

Connecting to the broader enforcement stack

The open-source proxy is the enforcement point. If you want centralized policy management across multiple agents, audit log aggregation, SIEM integration, or the compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001, those come from the Vaikora Control Plane, which is the commercial product hosted by Data443. The gateway itself, including MCP support, stays open source.

The two are designed to work together but are not required to. You can run the proxy standalone against a local policy file. When you are ready to manage policy centrally or push decisions into Microsoft Sentinel, SentinelOne, or CrowdStrike Falcon, the control plane connects without changing the proxy's position in the call path.

What to check next

Install `vaikora-guard-mcp` from PyPI and point it at an existing MCP server. Write one rule against a tool your agent already calls. Run the agent and read the audit log entry that comes back. The structure of that record is what every downstream integration consumes, so understanding it at the single-call level is the right place to start before scaling to multiple agents or adding the control plane.

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