VaikoraVaikora

Vaikora › Blog › Product Updates

Self-host a policy gate for AI agent tool calls: what the Vaikora Guard MCP proxy does and how to run it

Product Updates · Nov 17, 2026 · 3 min read

Installing a package from PyPI and running a self-hosted proxy is a faster path to AI agent policy enforcement than writing one from scratch. That is the practical reason `vaikora-guard-mcp` exists.

What problem the proxy solves

AI agents using the Model Context Protocol call external tools: file systems, databases, APIs, shell commands. Each call carries a request that the LLM constructed and the agent passes on without a second check. Nothing in the MCP specification itself enforces what that request is allowed to do. The proxy sits between the agent and those tool endpoints, reads each request before it reaches the tool, and applies a policy decision: allow, block, modify, or hold for human approval.

The decision happens before the tool executes. If the action is blocked, the tool never receives the call. The agent gets a structured refusal it can surface to the user or handle programmatically. No side effect occurs.

How the proxy works technically

`vaikora-guard-mcp` is a Python package, MIT licensed, published on PyPI. You pull it into your environment, configure policy, and point your MCP-compatible agent at it as a proxy endpoint instead of the target tool server.

At runtime the proxy:

That audit log entry exists regardless of outcome. Every allow is recorded with the same tamper-evident chain as every block. That matters for compliance audits where you need to show not just that bad calls were stopped, but that every call was checked.

Enforcement latency at P50 is 8ms, and P99 is 45ms. For most agent workflows, that is below the threshold where users notice a delay.

A concrete scenario

Say you have an AI coding agent with access to a file-write tool. A prompt injection in a code comment instructs the agent to overwrite a configuration file outside the project directory. Without a proxy, the LLM generates the tool call, the agent forwards it, and the file is overwritten before any log entry exists.

With `vaikora-guard-mcp` inline, the outbound tool-call request passes through the proxy. The prompt injection module flags the instruction as injected content. The policy rule for file-write calls outside permitted paths evaluates to block. The tool server never receives the request. The agent receives a structured refusal. The full decision, including the matched pattern and the policy rule that fired, is written to the audit chain.

Compliance presets and the control plane

The open-source proxy handles enforcement at the network layer. For teams that need centralized policy management across multiple agents, the Vaikora Control Plane is the commercial complement. It is hosted by Data443 and adds policy authoring, audit log search, approval workflows, and compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001.

The split matters: you can run the gateway yourself, inspect the code, and own the enforcement infrastructure. The control plane extends that into multi-agent governance without custom engineering.

Vaikora also ships prebuilt integrations with Microsoft Sentinel, CrowdStrike Falcon, and SentinelOne. High-severity blocks push into those platforms as threat indicators automatically, so your SOC sees AI agent risk alongside other signal without a separate pipeline.

What to check before you deploy

Before standing up the proxy in a production agent workflow:

The package is at `vaikora-guard-mcp` on PyPI. The source is MIT licensed, so you can read the inspection modules before you trust them. That is a reasonable thing to do before you put any proxy inline in a production system.

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