Vaikora › Blog › Product Updates

Vaikora Guard MCP is now open source: what the proxy does, how to run it, and what it enforces

Product Updates · Nov 5, 2026 · 3 min read

Tool calls in MCP-based agent workflows execute by default. There is no built-in mechanism to check whether a call should be allowed before it runs. Vaikora Guard MCP is a proxy that sits in that gap, and as of now it is published on PyPI under the MIT licence so you can inspect it, self-host it, and extend it.

What the proxy actually does

Vaikora Guard MCP sits inline between an AI agent and the tools it calls. When the agent issues a tool call, the proxy intercepts it, evaluates it against policy, and returns an allow, block, modify, or require-approval decision before execution. If the decision is block, the tool never runs. If it is require-approval, the call queues until a human responds.

This is different from logging after the fact. By the time a log entry exists, the tool has already run. The proxy enforces before execution.

The package is `vaikora-guard-mcp`, written in Python, and available on PyPI today.

Six content inspection modules out of the box

The gateway ships with six content inspection modules. These check things like whether a tool call contains credential patterns, whether it references paths or resources outside the agent's intended scope, and whether the payload matches policy rules you define. You configure policy in a file; the proxy applies it at runtime for every call.

Because the enforcement is deterministic, identical input always produces an identical decision. That matters for auditability: you can reason about what the proxy did and why, without chasing probabilistic model outputs.

Every decision is written to a SHA-256 hash-chained audit log. The chain is tamper-evident. If a decision record is altered or removed, the chain breaks.

Running it in under ten minutes

Install from PyPI:

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

Point your agent at the proxy instead of directly at your MCP server. The proxy acts as a pass-through for allowed calls and an enforcement point for everything else. No SDK migration is required. The proxy supports the major hosted model APIs, so if you are already running Claude, GPT-4o, or another provider, the adapter layer handles the translation.

A worked example: an agent is configured to call a file-read tool and an external HTTP tool. Without a policy layer, nothing stops it from reading `/etc/passwd` or posting to an unexpected endpoint. With Vaikora Guard MCP in front, you define a policy that restricts file reads to a specific directory and blocks outbound HTTP to any domain not on your allowlist. The proxy enforces both rules on every call, at P50 8ms and P99 45ms, so the latency impact on normal calls is negligible.

What the open-source layer covers and where the control plane fits

The MIT-licensed proxy is self-hosted and handles enforcement at the call level. The commercial Vaikora Control Plane, hosted by Data443, adds centralized policy management, team-level visibility across multiple agents and environments, SIEM integration (Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, AWS Security Hub), and compliance presets for SOC 2, HIPAA, GDPR, PCI-DSS, and ISO 27001.

If you are running a single agent in a dev environment, the open-source proxy is sufficient. If you are running multiple agents in production and need policy auditability across all of them with SIEM telemetry, that is where the control plane becomes the relevant piece.

What to check next

Start by installing `vaikora-guard-mcp` from PyPI and reviewing the default policy configuration. Map it against what your agent is actually permitted to do: which tools, which paths, which external endpoints. Where the defaults do not match your environment, edit the policy file before you route live traffic through the proxy.

If you want to see what calls are being intercepted before you enforce anything, run the proxy in audit-only mode first. The audit log gives you a complete record of every call and what decision would have been made, without blocking anything. That gives you a baseline before you turn enforcement on.

The source is on PyPI now. Read it, run it, and decide what your policy boundary should be before the next agent call executes without one.

More from the Vaikora blog