Vaikora › Blog › Product Updates

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

Product Updates · Oct 22, 2026 · 4 min read

Every tool call an AI agent makes is a decision point. Without a gate at that point, the agent executes whatever the model requests. Vaikora Guard MCP puts a policy check at that exact moment, before execution, and the code is now publicly available under the MIT license.

What the open-source release actually is

Vaikora Guard MCP is a self-hosted MCP-compatible proxy, written in Python and published on PyPI as `vaikora-guard-mcp`. It sits inline between an AI agent and any tool it calls. When the agent issues a tool call, the proxy evaluates it against configured policy. The result is one of four decisions: allow, block, modify, or require-approval. The tool call only proceeds if policy permits it.

The MIT license means you can inspect the code, modify it, deploy it in air-gapped environments, and integrate it into existing pipelines without a commercial dependency. The commercial product is the Vaikora Control Plane, hosted by Data443. The proxy itself is free to use and self-host.

Why "before execution" is the only position that matters

If a policy check runs after a tool call completes, it is forensics, not enforcement. An agent that has already written a file, sent an API request, or exfiltrated data cannot be stopped by a check that runs afterward. The proxy sits at the call boundary. The tool never receives the request unless policy clears it.

This is different from asking the model to evaluate its own behavior. Model-based self-evaluation is probabilistic: the same input can produce different outputs across runs. Vaikora Guard MCP applies deterministic policy. Identical input produces identical decisions, every time.

What gets checked on each call

The proxy ships six content inspection modules. Depending on how you configure it, a tool call can be evaluated for:

• sensitive data in parameters (PII, credentials, regulated content)
• whether the target tool or resource is on an allow-list
• whether the action class is permitted under the active policy preset
• whether the requesting agent or user context has the required approval

Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 ship with the proxy. You can activate a preset and have baseline enforcement running in minutes, then layer custom rules on top.

A concrete example

Say an agent is configured to query a database and summarize results. Without a gate, the agent can call any query the model constructs, including one that dumps a full customer table. With Vaikora Guard MCP inline, a rule blocking queries that match patterns for bulk record extraction stops that call before it reaches the database. The agent receives a policy-blocked response. The tool receives nothing. Every decision is written to a SHA-256 hash-chained, tamper-evident audit log, so you have a signed, ordered record of what was blocked and why.

That audit log is not optional. It writes on every decision, not just blocks. That matters for compliance reviews and incident investigations where you need to show exactly what an agent was and was not permitted to do over a given window.

Enforcement latency at production throughput

The objection to inline enforcement is always latency. The proxy runs at P50 8ms and P99 45ms. At 10,000+ actions per second, that is fast enough to sit in the critical path of a high-throughput agent pipeline without becoming the bottleneck. The gate does not require you to choose between safety and speed.

How it connects to the broader platform

The open-source proxy is the enforcement layer. If you want centralized policy management, multi-agent visibility, SIEM integration, or the full audit and reporting surface, that is the Vaikora Control Plane. The proxy connects to the control plane but does not require it. You can run policy locally from configuration files during development, then connect to the control plane for production governance.

Vaikora integrates natively with Microsoft Sentinel, SentinelOne, CrowdStrike Falcon, Microsoft Defender, and AWS Security Hub. If your security operations team monitors AI agent behavior in any of those platforms, enforcement signals from the proxy flow directly into your existing detection and response workflows.

What to check next

Install `vaikora-guard-mcp` from PyPI, point it at the tool your agent calls, and activate one of the compliance presets that matches your regulatory environment. Then review what gets blocked in the first hour. The audit log will show you what the model was trying to do before your policy stopped it.

More from the Vaikora blog