VaikoraVaikora

Vaikora › Blog › Product Updates

Why we open-sourced the policy gate for MCP tool calls, and what you can do with it today

Product Updates · Nov 21, 2026 · 3 min read

Every MCP-connected AI agent has a window between "tool call requested" and "tool call executed" where nothing checks whether that action should be allowed. The question we kept hearing from security engineers was simple: how do I put a policy decision in that window without forking the SDK or writing a custom proxy from scratch?

That question is why Vaikora Guard MCP is open source.

What the package actually is

`vaikora-guard-mcp` is a Python proxy, published on PyPI under the MIT license. It sits inline between an AI agent and any MCP-compatible tool server. When the agent requests a tool call, the proxy intercepts it, evaluates it against your policy file, and returns an allow, block, modify, or require-approval decision before execution starts.

If the decision is block, the tool call never reaches the server. If it is require-approval, execution waits. The agent does not proceed until the gate resolves.

Install it in one line:

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

Write a policy file. Point your MCP client at the proxy. That is the full setup path.

Why deterministic enforcement matters here

The enforcement model is intentional. Identical input always produces identical output. There is no probabilistic scoring, no "usually blocks" behavior. A policy rule either matches or it does not, and the same call gets the same answer every time.

That property matters for two reasons. First, it means you can test your policies in isolation before deploying them against production agent traffic. Second, it means every decision is auditable. Vaikora writes each enforcement outcome to a SHA-256 hash-chained, tamper-evident audit log, so you have an ordered record you can produce during an investigation or compliance review without reconstructing anything from memory.

The P50 enforcement latency is 8ms. P99 is 45ms. For most agent workflows, the gate adds no perceptible delay.

What you can enforce

A policy file can express rules against tool name, parameter values, data patterns in the request payload, and combinations of those. Some concrete examples of what teams configure:

The proxy ships with six content inspection modules, so pattern matching against PII, secrets, and known-bad content is built in rather than something you wire up yourself.

Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 are available, which gives teams with regulatory obligations a starting point that maps to control requirements rather than writing rules from first principles.

The self-hosted case and the enterprise case

The open-source proxy is fully self-hostable. You run it, you own the policy files, the audit log stays in your environment. There is no telemetry sent back, no call home. This matters for teams operating in air-gapped or high-compliance environments.

The commercial side is Vaikora, Data443's AI runtime control platform, which adds a hosted control plane, centralized policy management across multiple agents and environments, SIEM integrations with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub, and support for 12 LLM providers. The open-source proxy and the enterprise control plane use the same enforcement engine, so policy files you write against the proxy today carry forward.

What to do with this

If you are running MCP-connected agents now without a policy gate, the first concrete step is to install `vaikora-guard-mcp`, write a policy file that blocks the specific tool call patterns you would refuse if a human were reviewing them, and run it against your existing agent in a test environment. Observe what it would have blocked in the last session's traffic.

That exercise usually surfaces two or three tool call patterns that nobody intended to allow. Finding them before they execute in production is the point.

The source is on GitHub and the package is on PyPI. The README covers policy syntax and the inspection modules in detail.

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