MIT-licensed and published on PyPI as `vaikora-guard-mcp`, the Vaikora Guard MCP proxy is now open source. Any team running AI agents over the Model Context Protocol can self-host a policy enforcement layer that checks every tool call before it executes, without writing the enforcement logic themselves.
This post explains what the proxy does, how it works technically, and what the open-source release changes for teams trying to govern AI agents in real environments.
The problem this solves
AI agents make tool calls. Those calls touch file systems, APIs, databases, and external services. The tool call is the action, and once it executes, whatever damage it does is already done. A policy check that runs afterward is not a guardrail. It is an audit.
Most agent frameworks have no native enforcement layer. They rely on the agent’s own reasoning to decide whether a tool call is appropriate. That is not a security control. An agent that reasons incorrectly, gets manipulated through a prompt injection, or operates outside the scope its developers intended will still execute the call unless something external stops it.
Vaikora Guard MCP sits inline between the agent and the tools it calls. Every tool call passes through the proxy. The proxy applies a deterministic policy decision (allow, block, modify, or require approval) before the action executes. Identical input always produces an identical decision. That is the property that makes it auditable and testable.
What ships in the open-source package
The proxy is written in Python and available directly from PyPI:
“`
pip install vaikora-guard-mcp
“`
It is MCP-compatible, which means it works with any agent or framework that speaks the Model Context Protocol. No SDK migration is required for teams already using MCP-compatible tooling.
The package ships six content inspection modules. These cover the categories most relevant to AI agent risk: data exfiltration patterns, sensitive credential exposure, personally identifiable information in tool call payloads, policy scope violations, prompt injection attempts, and tool call frequency anomalies. Each module runs before execution, not as a background observer.
The proxy is self-hosted. Your tool call traffic does not leave your environment to be inspected. That matters for teams operating under HIPAA, PCI-DSS, or similar regulatory constraints.
A concrete scenario
Suppose an AI agent is tasked with summarizing customer support tickets and writing responses. Its tool set includes a file read tool, a database query tool, and an email send tool.
Without an enforcement layer, the agent can call any of those tools in any order, with any payload, as long as the underlying permissions allow it. If the agent is manipulated into exfiltrating customer records to an external address, the email send tool will execute. The agent’s reasoning did not catch it. The tool did not catch it.
With Vaikora Guard MCP in place, the tool call payload passes through the proxy before execution. The PII inspection module flags the customer record content. The policy for that agent profile blocks outbound data containing PII to unverified recipients. The call does not execute. The proxy writes the decision to a SHA-256 hash-chained audit log entry, with the full context of what was attempted, what policy matched, and what outcome was enforced.
The agent developer can review that log entry and understand exactly what happened. The compliance team has a tamper-evident record. No custom logging code was required.
Enforcement performance
A common concern with inline proxies is latency. The Vaikora gateway enforces policy at P50 8ms and P99 45ms, with throughput exceeding 10,000 actions per second. For most agent workflows, the enforcement overhead is not perceptible. For high-frequency agent loops, the throughput ceiling is well above what a typical agentic pipeline produces.
What the enterprise control plane adds
The open-source proxy handles enforcement at the tool call level. The Vaikora Control Plane, hosted by Data443, adds centralized policy management, team-level governance, compliance presets for SOC 2, HIPAA, GDPR, PCI-DSS, and ISO 27001, and native integrations with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.
The split is deliberate. Teams that want to run enforcement locally, without a commercial dependency, can do that with the MIT-licensed proxy. Teams that need policy management across multiple agent deployments, with SIEM integration and audit log export, have a path to the control plane without replacing the proxy they already deployed.
What to check next
If you are running AI agents in production and do not have an enforcement layer inline, start by mapping what tool calls those agents can make. List the tools, the permissions they carry, and what data they can touch. That inventory is the input to a policy definition exercise.
Then pull the `vaikora-guard-mcp` package, run it locally against your agent’s tool calls, and check what the inspection modules surface. The source is on PyPI and the license is MIT, so you can read the inspection logic, modify it, and deploy it without restriction.
The audit log the proxy produces will tell you more about what your agents are actually doing than any amount of reasoning-level analysis.