VaikoraVaikora

Vaikora › Blog › Product Updates

Deterministic policy before every agent tool call: a walkthrough of the Vaikora Guard MCP open-source release

Product Updates · Nov 19, 2026 · 3 min read

Every AI agent eventually calls a tool. That tool call might write to a database, query an API, send an email, or delete a file. Without a gate between the agent and the tool, the only control you have is the system prompt, and system prompts are not enforcement.

Vaikora Guard MCP is the open-source component that closes that gap. It is published on PyPI as `vaikora-guard-mcp`, written in Python, MIT-licensed, and self-hostable. The commercial Vaikora Control Plane sits on top, but the enforcement proxy itself is free to inspect, fork, and deploy.

What "before execution" actually means

When an MCP-compatible agent issues a tool call, the request travels through Vaikora Guard MCP before it reaches the tool server. The proxy evaluates the call against configured policy and returns one of four decisions: allow, block, modify, or require-approval. The tool does not execute until that decision is made.

This is a different model from post-hoc logging. A logging layer tells you what happened. An inline proxy decides what happens.

Identical input always produces an identical policy decision. That determinism matters for auditing: you can replay any decision and get the same outcome, which makes investigation straightforward and compliance reporting defensible.

What the package contains

Vaikora Guard MCP ships with six content inspection modules. These inspect the payload of a tool call for things like credential patterns, PII, prompt injection attempts, and policy violations before the call is forwarded.

It includes adapters for the major hosted model APIs and native MCP support, so you do not need to change how your agent is built. The proxy sits in the network path; the agent sees no difference.

Every decision is written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so the log is tamper-evident. If a record is altered, the chain breaks. That property is what makes the audit log useful for compliance presets including SOC 2, HIPAA, GDPR, PCI-DSS, and ISO 27001.

A concrete scenario

Say an AI agent is authorized to query a customer database and draft emails. Without a policy gate, there is nothing stopping the agent from issuing a tool call that extracts a full table and passes it to an email draft in the same session. Each individual action might be within the agent's stated scope, but the combination produces a data exposure.

With Vaikora Guard MCP inline, a policy rule can block any tool call that attempts to pass database query results directly into an outbound communication tool. The block happens before the email draft is created. The decision is logged, hash-chained, and visible to your security team through the Control Plane.

That is the enforcement pattern that post-execution logging cannot replicate. By the time a log entry appears, the email is already drafted.

Deployment path

Because the proxy is self-hosted and MIT-licensed, you can run it inside your own perimeter. No data leaves your environment to be evaluated. The commercial Control Plane, hosted by Data443, connects to the proxy to give you centralized policy management, the visual audit trail, and the SIEM integrations.

Vaikora integrates natively with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, Microsoft Defender, and AWS Security Hub. AI agent threats can be pushed into those platforms as enriched indicators, severity-mapped and actionable in your existing security operations workflow.

The enforcement latency is P50 8ms and P99 45ms, with throughput above 10,000 actions per second. Those numbers matter because enforcement that slows an agent to a crawl gets disabled. Policy that runs at agent speed gets left on.

What to check next

If you are evaluating the package, start at PyPI: search `vaikora-guard-mcp` and read the source before you deploy. The MIT license means you can inspect every inspection module and every policy evaluation path.

The questions worth answering before you go further: Which tool calls in your current agent deployment carry the highest data risk? Do you have a tamper-evident record of what your agents have already done? If the answer to the second question is no, that is the gap the audit chain is built to close.

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