When an AI agent makes a tool call, something has to decide whether that call is allowed. Without a gate inline, the answer defaults to "yes" every time, regardless of what the call does or what data it touches.
Vaikora Guard MCP is now published on PyPI under the MIT licence, installable with a single pip command, and self-hostable without any dependency on Data443’s commercial infrastructure. This post explains what the proxy actually does, how to wire it into an agent stack, and what you get from the open-source release versus the enterprise control plane.
What the proxy does, specifically
Vaikora Guard MCP sits between an MCP client (your agent or orchestrator) and an MCP server (the tool host). Every tool call the agent sends passes through the proxy before it reaches the server. The proxy runs the call through six content inspection modules, evaluates it against a deterministic policy, and issues one of four decisions: allow, block, modify, or require approval.
The key word is "before." The tool call does not execute until the proxy returns an allow decision. If the call is blocked, it never reaches the MCP server. If it requires approval, execution pauses until a human responds.
Because the policy is deterministic, identical inputs always produce identical decisions. There is no probabilistic scoring that returns a different result on a second run. That property matters for auditability: you can replay a sequence of calls and verify that the same policy would have produced the same outcome.
Installation and basic configuration
“`
pip install vaikora-guard-mcp
“`
The package is written in Python. Once installed, you point your MCP client at the proxy’s local endpoint instead of directly at the MCP server. The proxy forwards allowed calls upstream and handles the response path.
A minimal configuration file sets:
- The upstream MCP server address
- Which content inspection modules to enable (for example, PII detection, secret scanning, or prompt injection detection)
- The default action for calls that match no explicit rule (typically block or require-approval for production environments)
- Whether to write decisions to the local audit log
Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 ship with the package. Enabling a preset loads a baseline ruleset for that framework without requiring you to hand-write every rule from scratch.
A concrete scenario
An AI agent is tasked with querying a database and summarizing results. The agent constructs a SQL tool call that includes a raw customer email address in the WHERE clause, which it extracted from a previous step in the conversation.
Without a policy gate, that call executes. The database returns records scoped to that email, and the data flows back into the agent’s context, possibly including fields that were never meant to surface in an automated workflow.
With Vaikora Guard MCP inline and PII detection enabled, the proxy inspects the outbound tool call before it reaches the MCP server. The email address in the query body triggers the PII rule. The proxy blocks the call, logs the decision with a timestamp and the matched content class, and returns a structured error to the agent. The database never sees the request.
That block event writes to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so the record cannot be modified after the fact without breaking the chain. This is what makes the log usable in a compliance investigation, not just an operational log you can explain away.
Enforcement latency and throughput
The proxy adds P50 8ms and P99 45ms to a tool call. At 10,000+ actions per second, the inline inspection is not a meaningful bottleneck for most agent workloads. If your use case requires higher throughput, the self-hosted deployment gives you control over resource allocation.
What the enterprise control plane adds
The open-source proxy handles per-deployment policy enforcement locally. The Vaikora enterprise control plane adds centralized policy management across multiple deployments, SIEM integration (Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, AWS Security Hub), approval workflows with a UI, and the hash-chained audit log accessible across your organization rather than only on the host where the proxy runs.
Vaikora is available on AWS Marketplace and ships in the Azure Sentinel content hub. If you are already running Sentinel, the integration does not require custom engineering: high-severity decisions promote directly into Azure Security Center as native incidents.
What to check next
Pull the package and run it locally against a test MCP server. Enable one content inspection module, send a tool call that should trigger it, and verify that the audit log entry is written and the call is blocked. That takes under ten minutes and tells you whether the proxy behaves as expected in your environment before you write any production policy.
The source is on PyPI as `vaikora-guard-mcp`. Read the module configuration options before writing your first ruleset, specifically the default-action field. Leaving it set to "allow" during testing is fine. Leaving it there in production means every call without an explicit rule passes through unchecked.