VaikoraVaikora

Vaikora › Blog › Product Updates

Why we published vaikora-guard-mcp under MIT, and what the source tells you about how enforcement works

Product Updates · Nov 22, 2026 · 4 min read

Publishing the proxy as open source was a deliberate choice, not a default. When policy runs between an agent and its tools, the people responsible for security need to be able to read the code, run it in their own environment, and trust that the decision logic is what it claims to be. A closed binary enforcing access control is a contradiction in terms.

Here is what the release actually contains, what it does at runtime, and how to think about where it fits in a production deployment.

What the package is

vaikora-guard-mcp is a Python package, MIT licensed, published on PyPI under that name. It is a self-hosted MCP-compatible proxy that sits inline between an AI agent and the LLM or tools it calls. Every tool call passes through it before execution. The proxy makes an allow, block, modify, or require-approval decision on each call, then passes the result downstream or stops it entirely.

The package ships six content inspection modules and adapters for the major hosted model APIs. Because it implements the MCP protocol natively, it works with any MCP-compatible agent without an SDK migration. Data443 also supports 12 LLM providers through the same interface.

Identical input always produces an identical policy decision. That determinism matters for auditability: if a call was blocked at 14:32 UTC, the same call submitted five minutes later returns the same result, and both decisions appear in the audit log.

What happens when a tool call arrives

The sequence is worth spelling out concretely. Suppose an agent calls a `write_file` tool with a path argument of `/etc/passwd`. The proxy receives that call before it reaches the tool. It runs the call through the configured inspection modules: checking the tool name against the allow-list, inspecting arguments for policy violations, evaluating any context rules the operator has written.

If a rule matches a block condition, the call does not execute. The proxy returns a structured refusal to the agent. The decision is written to a SHA-256 hash-chained audit log, tamper-evident and ordered. If the operator has configured require-approval for a sensitivity tier, the call is held and routed for human review before anything runs.

Enforcement latency at P50 is 8ms, P99 is 45ms, and the proxy handles more than 10,000 actions per second. Those numbers matter in agentic workflows where many tool calls happen in rapid succession.

What you can verify by reading the source

The MIT license means you can read every line. Specifically worth reviewing:

If your compliance team needs to demonstrate that policy decisions are deterministic and tamper-evident, the audit log structure is the place to start. Each record links to the previous record's hash, so any post-hoc modification breaks the chain.

Compliance presets and the control plane

The open-source proxy enforces policy, but policy has to come from somewhere. vaikora-guard-mcp can run with rules you write directly in configuration. The Vaikora Control Plane, hosted by Data443, adds centralized rule management, the compliance presets (SOC 2, HIPAA, GDPR, PCI-DSS, ISO 27001), and the integrations that push decisions into Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.

The proxy is also available on AWS Marketplace as a SaaS subscription, and the Vaikora-AzureSecurityCenter solution ships in the Azure Sentinel content hub for teams already running in that environment.

For teams that want full control, self-hosting the proxy with locally managed rules is a supported path. For teams that need centralized governance across multiple agents or environments, the control plane handles distribution and audit aggregation.

What to check next

If you are evaluating the package, the most useful starting point is to run a local instance with one rule and submit a test tool call that should be blocked. Watch what comes back to the agent and read the audit record that gets written. The structure of that record is what your SOC or compliance team will work with in production, so understanding it before you deploy at scale saves significant rework later.

The PyPI page and the source repository both contain the full rule schema. Write a rule against a realistic tool call from your own environment, not a toy example, and you will find the edge cases that matter for your use case before they matter in production.

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