Vaikora › Blog › Product Updates

How to run policy checks on every AI agent tool call: an open-source starting point with Vaikora Guard MCP

Product Updates · Nov 10, 2026 · 4 min read

Building an AI agent that calls external tools is straightforward. Building one where every tool call passes a policy check before it executes is not, unless you have an enforcement layer sitting inline.

That is exactly what Vaikora Guard MCP is. It is an open-source, MIT-licensed proxy written in Python, published on PyPI as `vaikora-guard-mcp`, and designed to sit between an MCP-compatible AI agent and the tools it calls. Every action goes through a policy decision first. If the action is blocked, the tool never runs.

What "inline" actually means here

A lot of enforcement approaches work by logging what an agent did, then alerting afterward. Vaikora Guard MCP works differently. The proxy intercepts the tool call before it reaches the tool. The agent sends a request, the proxy evaluates it against your policy, and one of four decisions comes back: allow, block, modify, or require-approval. Only an allow (or an approved request) moves forward to execution.

This is deterministic. The same input, evaluated against the same policy, always produces the same decision. There is no probabilistic model deciding whether a tool call "looks risky." The policy either permits it or it does not.

Enforcement latency at P50 is 8ms and at P99 is 45ms. For most agent workflows, that is not noticeable.

How to get it running

```
pip install vaikora-guard-mcp
```

From there, you configure your MCP server to route tool calls through the proxy endpoint rather than directly to the tool. No SDK migration is required. The proxy ships adapters for major hosted model APIs and handles the MCP protocol layer, so the agent code itself does not change.

A worked example: suppose you have an agent that can call a file-write tool. Without a gate, the agent can write to any path the host OS permits. With Vaikora Guard MCP in front of it, you define a policy that blocks writes to paths outside an allowed directory. If the agent constructs a write request pointing somewhere else, the proxy blocks it before the filesystem is touched. The agent receives a denial response. No write happens.

You can apply the same pattern to database queries, API calls, shell execution, or any other tool an agent can invoke through MCP.

What ships in the proxy

The gateway includes six content inspection modules that evaluate the payload of a tool call, not just its type. That means you can write policy based on what data is moving, not only which tool is being called.

The proxy writes every decision to a SHA-256 hash-chained, tamper-evident audit log. Each entry records what was requested, what policy evaluated it, and what decision was returned. Because the chain is hash-linked, you can verify after the fact that the log was not modified. That matters for compliance teams doing investigations after an incident, or for auditors who need to confirm that a specific action was blocked.

Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 are included, so you are not writing policy from scratch for regulated environments.

The difference between the open-source proxy and the enterprise control plane

Vaikora Guard MCP is the enforcement point. It is self-hostable, open source, and free to use under MIT. The Vaikora Control Plane, hosted by Data443, adds centralized policy management, real-time telemetry across multiple agents, and native integrations with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.

For teams running a single agent in a controlled environment, the open-source proxy may be enough to start. For teams running many agents across different environments, or who need SOC-visible telemetry, the control plane is where the operational picture comes together.

Vaikora also ships on AWS Marketplace as a SaaS subscription, which means it can draw against an existing EDP commitment if that is relevant to your procurement path.

Where to start

If you want to test the enforcement path before committing to a deployment, the fastest route is to `pip install vaikora-guard-mcp` and point a test agent at the proxy with a simple allow/block policy on a specific tool type. Run a few calls, check the audit log output, and confirm you can see the decision record for each one. That will tell you whether the enforcement model fits your agent architecture before you invest time in production policy design.

The source is on PyPI and the license is MIT, so there is nothing to sign before you read the code.

More from the Vaikora blog