Vaikora › Blog › Product Updates
MIT-licensed, self-hostable, 8ms median latency: what Vaikora Guard MCP actually does at the call layer
Shipping an AI agent without a policy gate on its tool calls means every tool executes at the agent's discretion. The only question is whether you find out about a bad call before or after the damage is done. Vaikora Guard MCP is built to answer that before it executes.
What the open-source release actually is
Vaikora Guard MCP is a Python package, published on PyPI as `vaikora-guard-mcp`, MIT licensed and self-hostable. It sits inline as an MCP-compatible proxy between an AI agent and the tools or LLMs it calls. Before any tool call executes, the proxy evaluates it against policy and returns one of four decisions: allow, block, modify, or require-approval.
The package ships six content inspection modules and adapters for the major hosted model APIs. It supports 12 LLM providers with no SDK migration required on your side.
Deterministic enforcement is the key property here. Identical input always produces identical output. There is no probabilistic scoring that lets an edge case slip through on the second attempt. The gate either passes the call or it does not.
What happens at runtime
Consider a concrete scenario. An AI agent is asked to retrieve a customer record and then write the result to an external webhook. Without a policy gate, both steps execute sequentially with no review. With Vaikora Guard MCP in the path:
- The agent constructs the tool call to fetch the customer record.
- The proxy intercepts it, checks it against the active policy (say, a HIPAA preset that restricts what data fields can flow outside the classification boundary).
- If the destination webhook is not on an approved list, the proxy blocks the second call before the data leaves.
That second step never executes. The agent receives a structured block response and can surface it to the user or escalate for approval. The whole exchange at P50 latency is 8ms. At P99 it is 45ms.
The audit chain
Every decision the proxy makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry references the previous one, so any deletion or modification of a record breaks the chain. That property matters for compliance reporting: you can demonstrate to an auditor exactly which tool calls were allowed, which were blocked, and which required approval, in sequence, with no gaps.
Compliance presets ship for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. You can apply one of these presets or build a custom policy on top of the same enforcement layer.
How the open-source layer relates to the control plane
The proxy itself handles runtime enforcement. The Vaikora Control Plane, hosted by Data443, is the commercial layer that adds centralized policy management, cross-agent visibility, and the integrations that push decisions into your existing security operations tooling.
Those integrations include Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. High-severity AI agent threats surface in SentinelOne as Threat Intelligence indicators mapped at severity 100 for Critical. In Azure, they promote directly into Microsoft Defender for Cloud as native incidents with no custom engineering.
The gateway is also available as a SaaS subscription on AWS Marketplace, with support for private offers and EDP draw-down.
What to check before you deploy
Before you drop the proxy into a production agent workflow, verify three things:
- Policy scope. Which tool categories should be blocked by default and which require approval rather than a hard block? Define that before you run traffic through the proxy, not after.
- Approval routing. The require-approval decision needs a destination. Decide whether that is a human-in-the-loop webhook, a ticketing system, or a Sentinel alert before the first approval-required call fires.
- Audit log retention. The hash-chained log is only as useful as your retention policy. Set a retention window that matches your compliance requirement on day one.
The package is on PyPI. Install it, point it at your MCP tool server, and run a tool call through it in a test environment. The block and audit behavior is visible immediately, before you commit to any configuration changes 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
Vaikora