Vaikora › Blog › Product Updates

vaikora-guard-mcp is MIT-licensed and on PyPI. Here is what it enforces and how to run it.

Product Updates · Oct 22, 2026 · 4 min read

Most MCP servers pass tool calls straight through to the model or the downstream service. `vaikora-guard-mcp` sits in front of that and refuses, modifies, or holds calls for approval before they touch anything. The code is on PyPI under the MIT licence, so you can read every enforcement decision it makes.

What the problem looks like without a policy gate

An AI agent connected to an MCP server can call tools in sequence, often without any human in the loop. If the agent is compromised, misconfigured, or simply given ambiguous instructions, it can read files it should not read, call APIs with credentials it should not hold, or exfiltrate data in a single tool call. Standard MCP has no built-in policy layer. The server accepts or rejects calls based on schema validation, not on your organization's rules about what agents are allowed to do.

This is the gap `vaikora-guard-mcp` addresses.

What the proxy actually does

`vaikora-guard-mcp` sits inline between your MCP client (the agent) and the downstream tool or LLM. Every tool call passes through six content inspection modules. The proxy evaluates the call against your configured policy and produces one of four deterministic decisions: allow, block, modify, or require-approval.

Deterministic is the operative word. Identical input produces an identical decision. There is no probabilistic scoring that lets a slightly reworded prompt slip through. The policy is a set of rules; the enforcement is a gate.

Every decision is written to a SHA-256 hash-chained, tamper-evident audit log. You can verify that the log has not been altered and that every tool call your agent made was evaluated. That property matters for compliance: SOC 2, HIPAA, GDPR, PCI DSS, and ISO 27001 presets ship out of the box.

Enforcement latency sits at P50 8ms and P99 45ms. At 10,000 or more actions per second, this is not a throttle on agent throughput for most workloads.

How to run it

Install from PyPI:

```
pip install vaikora-guard-mcp
```

The package is written in Python. Because it ships as a standard MCP-compatible proxy, it works with any MCP client without an SDK migration. You configure policy in the control plane or via local config, point your MCP client at the proxy endpoint, and the proxy forwards permitted calls onward.

A concrete example: you have an agent that can call a file-read tool, a SQL query tool, and an external HTTP tool. You want to allow file reads from a specific directory, block all external HTTP calls, and require human approval for any SQL query that touches a table named `payments`. You define those three rules. The proxy enforces them before any tool executes. The agent never sees a response from the SQL tool unless a human approves the call.

That approval workflow is auditable. The log entry for the held call includes the tool name, the arguments, the timestamp, the policy rule that triggered the hold, and the hash tying it to the previous entry in the chain.

What open-source means here specifically

The MIT licence means you can read the enforcement logic, fork it, and run it in an air-gapped environment. `vaikora-guard-mcp` is self-hosted. There is no telemetry leaving your network unless you connect it to the Vaikora Control Plane, which is the commercial product hosted by Data443.

Open-sourcing the proxy without the control plane is a deliberate split: the enforcement runtime is inspectable and self-contained; the policy management, SIEM integrations, and enterprise audit features are in the hosted tier. For teams that want to wire the proxy into Microsoft Sentinel, CrowdStrike Falcon, or SentinelOne, that happens through the control plane.

Where to go from here

• Pull the source on PyPI (`vaikora-guard-mcp`) and read the content inspection modules before you deploy.
• Check which of your MCP-connected agents have tools that touch sensitive data stores, external APIs, or credentials. Those are the tool calls that need a policy gate first.
• If you need the compliance presets or the SIEM integrations, the Vaikora AWS Marketplace listing and the Azure Sentinel content hub both carry the full control plane.

The practical first step is auditing which tool calls your agents currently make with no policy evaluation at all. That list is usually longer than expected.

More from the Vaikora blog