VaikoraVaikora

Vaikora › Blog › Product Updates

vaikora-guard-mcp: an open-source MCP proxy that blocks policy violations before tool calls execute

Product Updates · Nov 18, 2026 · 4 min read

When an AI agent violates policy, it usually means a tool call already ran. The file was read, the API was hit, the message was sent. Blocking after the fact is not enforcement, it is cleanup. vaikora-guard-mcp is a small, MIT-licensed Python package that sits between an MCP client and the tools it calls, and applies a policy decision before execution, not after.

What it is and what it does

vaikora-guard-mcp is an MCP-compatible proxy. When an agent issues a tool call, the request passes through the proxy first. The proxy evaluates the call against configured policy and returns one of four decisions: allow, block, modify, or require-approval. If the decision is block, the tool never runs. The agent receives a structured response describing why.

The package is published on PyPI as `vaikora-guard-mcp`, written in Python, and self-hosted. No data leaves your environment. The source is available under the MIT licence, so you can read exactly what the proxy does to each request, modify it, or contribute to it.

This matters because MCP is becoming the standard wire protocol for agentic tool use. A proxy at that layer intercepts every tool call regardless of which model or agent framework is making the request.

How the enforcement decision works

The proxy checks each incoming tool call against a deterministic policy definition. Identical input produces an identical decision every time. There is no probabilistic score, no threshold to tune, no model to drift. If a call matches a block rule, it is blocked. Period.

This is different from how most AI safety controls work today. Many systems observe what an agent did, classify it post-hoc, and alert. The proxy acts before the tool call reaches the tool. The window between "agent decides to act" and "tool executes the action" is where enforcement has to happen, and that is exactly where the proxy lives.

Consider a concrete example. An agent is configured to query a customer database and summarize records. A prompt injection in one of those records instructs the agent to also export the full table to an external endpoint. Without a policy gate, the export call runs. With the proxy enforcing a rule that blocks any tool call writing data to an external URL not on an approved list, the call is stopped before execution. The agent logs a blocked action. No data leaves.

What gets logged

Every decision the proxy makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry contains the tool name, the full call parameters, the policy that matched, the decision, and a timestamp. Because entries are chained, any modification to a past record breaks the chain and is detectable.

This is relevant for compliance teams. SOC 2, HIPAA, GDPR, PCI-DSS, and ISO 27001 all have requirements around demonstrating that controls were in place and operating. A hash-chained log of every enforcement decision is a verifiable record, not just a claim that controls existed.

The relationship to the Vaikora Control Plane

The open-source proxy enforces policy locally using rules you define and manage yourself. The commercial product, the Vaikora Control Plane hosted by Data443, adds centralized policy management, pre-built compliance presets for the frameworks listed above, integration with SIEM platforms including Microsoft Sentinel and CrowdStrike Falcon, and the full enterprise audit chain.

For teams that want to start with the open-source proxy, the path to the control plane does not require changing your agent code or switching LLM providers. Vaikora supports 12 LLM providers with no SDK migration. The proxy layer stays in place; the policy management and telemetry destination change.

What to check before you deploy

Before you install the proxy, answer three questions:

Install `vaikora-guard-mcp` from PyPI, review the policy schema in the repository, and write a rule for one high-risk tool call. Run it in a test environment and verify that the blocked call produces a log entry and that the tool never received the request. That sequence, install, write one rule, verify the block, is the right starting point before expanding policy coverage across a full agent deployment.

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