MIT license, Python, published on PyPI as `vaikora-guard-mcp`. That is the entire distribution story for Vaikora Guard MCP. If your team is running AI agents today and wondering where policy enforcement actually goes, the answer is: inline, before the tool call executes, and now available to install in one command.
The Problem This Solves
AI agents do not wait for approval. They call tools, query databases, write files, and invoke APIs based on whatever the LLM decides to do next. Standard access controls such as Azure RBAC, AWS IAM, and Snowflake roles set a permission boundary, but they do not evaluate the content of a call or apply context-sensitive policy at runtime. An agent with read access to a customer database has that access regardless of what it is about to do with the data.
Vaikora Guard MCP sits inline between the agent and whatever tool or model it is calling. Before any action executes, it applies a deterministic policy check and returns one of four decisions: allow, block, modify, or require-approval. The same input always produces the same decision. That is not a probabilistic filter. It is a gate.
What "Open Source" Actually Means Here
The `vaikora-guard-mcp` package is MIT licensed and self-hostable. Your team can read the code, audit the policy logic, run it in your own infrastructure, and contribute changes. There is no call home, no telemetry sent to Data443 by default, and no SDK migration required to connect it to the LLMs you are already using. It supports 12 LLM providers as-is.
The enforcement point ships as an MCP-compatible proxy. For teams already building on the Model Context Protocol, that means dropping Vaikora Guard MCP into your existing architecture without redesigning how your agents communicate with tools.
The commercial product is the Vaikora Control Plane, hosted by Data443. That is where you manage policies centrally, review audit logs, and handle approval workflows across multiple agents. The open-source gateway enforces those policies at runtime. You can run the gateway without the control plane if your use case is simpler.
A Concrete Setup Scenario
Say you have a customer-support agent that can query a CRM, send emails, and write to a ticketing system. Without a runtime gate, any prompt injection or unexpected model behavior could cause the agent to exfiltrate contact records or send unauthorized emails. Standard role-based access does not catch that because the agent has legitimate access.
With Vaikora Guard MCP installed as the proxy:
- Every tool call passes through the enforcement point before it reaches the CRM, email API, or ticketing system.
- A policy rule blocks any outbound payload containing PII fields that were not part of the original user request context.
- A separate rule routes any action touching payment fields to a require-approval queue.
- Every decision, allow or block, writes to a SHA-256 hash-chained audit log. The chain is tamper-evident, so you can produce an ordered record of what the agent did and when for any compliance review.
That audit log matters separately from enforcement. When a compliance auditor or incident responder asks what an agent did between 2 p.m. and 4 p.m. on a given day, the answer is in the log, not reconstructed from inference.
Performance Characteristics
Enforcement latency runs at P50 8ms and P99 45ms. The gateway handles 10,000 or more actions per second. For most agent architectures, those numbers mean the enforcement layer is not the bottleneck. You are not trading policy control for throughput.
Detection accuracy in testing is 99.9%. That figure covers the content inspection modules that evaluate whether a given tool call matches a policy rule. Vaikora ships six content inspection modules in the gateway, covering prompt content, data type patterns, destination checks, and similar signals.
Compliance Presets
If your environment has specific regulatory requirements, Vaikora ships presets for SOC 2, HIPAA, GDPR, PCI DSS, and ISO 27001. A preset is a starting policy configuration that maps agent actions to the controls those frameworks require. You can use a preset as-is or extend it with custom rules.
What to Check Next
Install the package from PyPI and review the policy configuration format before you deploy anything to production. Understand which tool calls your agents are making today, specifically which ones touch sensitive data or external systems. Map those against the four decision types (allow, block, modify, require-approval) and decide where you need a hard block versus a human-in-the-loop approval step.
The source code is public. Reading it is a reasonable first step before trusting any inline enforcement layer with your agent traffic.