Home | Blog | Vaikora Guard MCP: An Open-Source Proxy That Enforces Policy on AI Tool Calls Before They Run

Vaikora Guard MCP: An Open-Source Proxy That Enforces Policy on AI Tool Calls Before They Run

Every MCP-compatible AI agent can call tools. Nothing in the protocol itself stops an agent from deleting a file, exfiltrating data, or hitting an external endpoint it should never reach. The gap is a policy layer that sits inline, checks each call, and either allows it or stops it before execution begins.

Vaikora Guard MCP is that layer. It is MIT-licensed, written in Python, published on PyPI as `vaikora-guard-mcp`, and self-hostable. This post covers what it does, how it works technically, and when you would want it.

What the proxy actually does

Vaikora Guard MCP sits between your AI agent and the tools it calls. When the agent issues a tool call, the proxy intercepts it, evaluates it against your configured policy, and returns one of four decisions: allow, block, modify, or require-approval. The underlying tool only runs if the proxy allows it.

The decisions are deterministic. Identical input produces identical output every time, which matters for auditability and debugging. Every decision is written to a SHA-256 hash-chained, tamper-evident audit log, so you have an ordered record of what was checked, what was decided, and when.

Enforcement latency at P50 is 8ms and at P99 is 45ms, so the overhead added to any single tool call is small enough to leave unchanged in almost every real-world workflow.

The enforcement model

The proxy ships six content inspection modules. These cover the range of things you typically want to check: whether a tool call contains sensitive data patterns, whether a destination is on an allowlist or blocklist, whether the action type is permitted under current policy, and similar.

Policy is configured declaratively. You define what is permitted, what is blocked outright, what requires a human approval step, and what should be modified before passing through. The proxy enforces that configuration at runtime, not after the fact.

A worked example: suppose an agent is authorized to query a database but should never write to it. You configure a policy rule that blocks any tool call whose action maps to a write operation on that data source. The agent issues a write call. The proxy intercepts it, matches it against the write-block rule, and returns a block decision. The database never receives the call. The block is logged with the full tool call payload, the matched rule, the timestamp, and the hash of the preceding log entry.

This is different from monitoring. Monitoring tells you after execution that something went wrong. The proxy prevents execution when policy says no.

Installing and running it

“`
pip install vaikora-guard-mcp
“`

The package is on PyPI. Because it is MCP-compatible, it plugs into any agent framework that supports the Model Context Protocol without requiring an SDK migration. Data443 supports 12 LLM providers in the same architecture.

Self-hosting means the proxy runs in your own infrastructure. No tool call payload leaves your environment to be evaluated. That matters for HIPAA, PCI-DSS, and GDPR deployments, and the proxy ships compliance presets for all of them, including SOC 2 and ISO 27001.

When you need this and when you do not

If your agents only call tools in a fully sandboxed, read-only environment with no access to sensitive data or external systems, the operational risk is low and the proxy is optional. Add it anyway if you want an audit record.

If your agents call tools that touch customer data, write to any persistent store, make external HTTP calls, or operate under any compliance requirement, the proxy is not optional. The compliance question is not whether something went wrong; it is whether you can prove what happened and when.

The require-approval decision path is specifically for high-risk actions you want a human to review before they run. The agent submits the call, the proxy holds it, a human approves or rejects, and the proxy either releases or blocks. That loop happens inline with the full audit chain intact.

The open-source layer and the enterprise layer

The MIT-licensed proxy handles enforcement. The enterprise control plane, hosted by Data443, adds centralized policy management, integration with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub, and the management surface for teams running multiple agents across multiple environments.

You can run the open-source proxy standalone. For teams that need SIEM integration or cross-agent governance, the control plane is the addition.

What to check next

Start with the PyPI package and run it against a single agent in a test environment. Write one policy rule that blocks a call type you would never want in production, trigger it deliberately, and inspect the audit log entry. That sequence tells you more about the enforcement model than any documentation can. From there, map your actual tool call inventory against the action types the proxy can inspect and decide which calls need allow, block, modify, or require-approval treatment before you move to production.