Vaikora › Blog › Product Updates

We open-sourced an MCP proxy that checks tool calls against policy before they run. Here is how it works.

Product Updates · Nov 9, 2026 · 4 min read

Policy enforcement for AI agents has a placement problem. Most teams add controls around the agent, not inside the call path. By the time a log entry appears or an alert fires, the tool has already run, the file is already written, and the downstream action is already in motion.

Vaikora Guard MCP is a small, self-hosted proxy that sits between an MCP-compatible AI agent and the tools it calls. Every tool call passes through it. The proxy applies a deterministic policy check and returns an allow, block, modify, or require-approval decision before execution starts. It is open source, MIT-licensed, written in Python, and published on PyPI as `vaikora-guard-mcp`.

What "inline" means here

The proxy is not a sidecar that observes after the fact. It is in the call path. The agent issues a tool call. The proxy intercepts it, evaluates it against policy, and either permits the call, blocks it, rewrites the parameters, or holds it for a human to approve. Only after that decision does execution proceed.

This matters because the alternative, checking logs after the fact, tells you what happened. Inline enforcement determines what is allowed to happen.

Identical input always produces an identical policy decision. There is no probabilistic scoring that lets the same call pass on Monday and fail on Thursday. That determinism is what makes the control auditable and what makes investigations straightforward.

What the proxy actually inspects

Vaikora Guard MCP ships with six content inspection modules. These check things like:

• Whether the tool call targets a resource the agent is permitted to access under current policy
• Whether the payload contains content that matches defined sensitivity patterns, such as credential material, PII, or regulated data categories
• Whether the action type (read, write, delete, external call) is permitted for the agent identity making the request
• Whether a human approval step is required before the call proceeds

The proxy adapts to the major hosted model APIs and supports MCP natively, so it fits into an existing agent setup without an SDK migration.

Latency at enforcement scale

A common objection to inline policy gates is that they add unacceptable latency. The numbers here are specific: P50 enforcement latency is 8ms, P99 is 45ms, and the proxy handles 10,000 or more actions per second. For most agent workflows, that is not a bottleneck. For agentic pipelines running at high throughput, the throughput figure is more relevant than the latency figure, and both hold.

The audit chain

Every decision the proxy makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each record is signed into the chain in order. If a record is altered or deleted, the chain breaks. This gives you an ordered, verifiable record of every tool call decision, what was submitted, what policy was applied, and what the outcome was.

For compliance purposes, this is the difference between "we believe our agent behaved correctly" and "here is a signed record of every decision it was subject to, in order, that you can verify independently." Vaikora ships compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001.

Connecting to the rest of your security stack

The self-hosted proxy is the open-source component. The Vaikora Control Plane, the commercial product hosted by Data443, is where you manage policy centrally, review approval queues, and push enforcement signals to your existing security tooling.

The proxy integrates natively with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. High-severity AI agent actions appear as threat intelligence indicators in SentinelOne (severity-mapped: Critical=100, High=75, Medium=50, Low=25) and as native incidents in Azure Security Center and Microsoft Defender for Cloud, with no custom engineering on your end.

Vaikora also adds a runtime gate on top of Azure RBAC, AWS IAM, and Snowflake roles rather than replacing them. If the agent has IAM permission but policy says no, the call is blocked before it reaches the target.

What to do next

If you want to see the proxy before committing to anything, the code is on PyPI: `pip install vaikora-guard-mcp`. Review the six inspection modules and the policy schema to understand what it can and cannot check in your specific tool call surface. Then map your highest-risk tool types (write operations, external calls, anything touching credentials or regulated data) against the policy options available, and decide where you want allow, block, modify, or require-approval applied.

More from the Vaikora blog