Vaikora › Blog › Product Updates
What vaikora-guard-mcp actually does when a tool call arrives: a walkthrough of the enforcement path
Tool calls are where AI agent risk becomes real. The agent has already decided what to do. The tool call is the moment the decision turns into an action. If your policy check runs after that point, you are auditing outcomes, not controlling them.
vaikora-guard-mcp is an open-source, MIT-licensed MCP-compatible proxy that sits in that exact position. It is published on PyPI as `vaikora-guard-mcp`, written in Python, and self-hostable. The enforcement path is worth understanding in detail because the design choices are specific.
What happens when a tool call arrives
An AI agent constructs a tool call and sends it to the MCP server. vaikora-guard-mcp intercepts that call before it reaches the tool. It runs the call through a set of content inspection modules, checks the result against your policy, and returns one of four deterministic decisions: allow, block, modify, or require-approval.
Identical input always produces identical output. There is no probabilistic scoring, no confidence threshold to tune. The policy either matches or it does not. That determinism matters for auditability: you can replay any decision and get the same result.
Every decision is written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so the record is tamper-evident and ordered. If a tool call was blocked at 14:32:07, that fact is in the chain, and the chain cannot be edited to remove it.
What "inline" means in practice
Consider a scenario: an agent is tasked with summarizing customer support tickets. The tool it calls has read access to a support database. The agent constructs a query that happens to include a parameter pulling records outside the intended scope, perhaps a customer ID range that covers accounts the agent should not touch.
Without inline enforcement, that query runs. The data comes back. The scope violation happened. You may detect it later in a log review or not at all.
With vaikora-guard-mcp in the path, the call arrives at the proxy. A content inspection module checks the parameters against the policy for that tool, specifically which customer ID ranges are permitted for this agent identity. The call exceeds the allowed range. The proxy returns a block decision. The query never reaches the database. The block is written to the audit chain with the full call payload, the matched policy rule, and a timestamp.
The agent receives a structured rejection it can handle in code. The enforcement happened before execution.
Configuration and inspection modules
The gateway ships six content inspection modules. You configure which modules apply to which tools and which agent identities. A tool that handles PII gets different inspection than a tool that reads from a public API.
Compliance presets ship for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. A preset is a starting policy configuration, not a certification. It gives you a documented baseline you can adjust, and the audit log records every decision against that baseline.
The proxy supports the major hosted model APIs, so you are not locked to a single provider. Adapters handle protocol translation; your policy layer stays consistent regardless of which LLM the agent uses.
Self-hosted versus the control plane
vaikora-guard-mcp is the open-source component. You run it yourself. The policy configuration, the inspection modules, and the audit log are all local to your infrastructure. There is no data leaving your environment.
Data443's Vaikora Control Plane is the commercial product, hosted by Data443. It adds centralized policy management across multiple proxies, enterprise integrations with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub, and the ability to push high-severity AI agent threats as native incidents into those platforms without custom engineering.
The two are designed to work together, but the proxy functions independently. For teams that want to start with the open-source layer and evaluate what they need beyond it, that is a reasonable path.
What to check before you deploy
Before standing up vaikora-guard-mcp, map which tool calls your agents actually make in production. The proxy enforces policy against calls it sees. If an agent has a path to a tool that bypasses the proxy, that path is outside your control boundary.
Specifically:
- Identify every tool registered in your MCP server and the agent identities that can call each one.
- Decide which tools handle sensitive data, external network calls, or write operations. Those are your highest-priority enforcement points.
- Write a policy rule for each of those tools before you go live. A proxy with no policy configured allows everything.
- Confirm the audit log is being persisted to a location your security team can query. The hash chain is only useful if someone is reading it.
The PyPI package is `vaikora-guard-mcp`. The source is MIT-licensed. Start there, read the content inspection module configuration, and map it against your actual agent tool surface before you run it in front of anything sensitive.
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