Every tool call an AI agent makes is a decision point. The agent decides to call a function, pass arguments, and act on the result. Without a gate at that point, the policy question "should this action be allowed?" never gets asked until after something goes wrong.
Vaikora Guard MCP is a small, MIT-licensed Python package that sits inline between an agent and its tools. It checks each tool call against a policy before the call executes, writes the decision to an audit log, and either allows, blocks, or modifies the call. The source is on PyPI as `vaikora-guard-mcp`. You can read it, run it on your own infrastructure, and modify it without signing anything.
What "inline" actually means here
The proxy receives the tool call request from the agent, evaluates it against the configured policy, and returns a verdict. Allow means the call proceeds. Block means it does not, and the agent receives a structured refusal. Modify means the proxy rewrites the arguments before passing them on. Require-approval pauses execution and waits for a human decision.
This happens before execution, not after. A detection-after-the-fact approach catches the call in a log. An inline approach catches it before the tool runs.
Enforcement latency is P50 8ms, P99 45ms, so the gate does not add meaningful delay to agent workflows running normal tool volumes. At the high end, the gateway handles 10,000+ actions per second.
How the policy model works
Policy decisions in Vaikora are deterministic. The same input always produces the same output. That matters for two reasons.
First, it makes the system auditable. Every decision is signed into a SHA-256 hash-chained, tamper-evident audit log. If a regulator or incident investigator asks what happened during a specific agent session, the log gives an ordered, verifiable record of every call, verdict, and reason.
Second, it makes policy testable. You can write a tool call, run it through the proxy, and confirm the verdict before you deploy. That is not possible with probabilistic LLM-based guardrail approaches, where the same prompt can produce different verdicts on different runs.
Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 ship out of the box. These are starting configurations, not complete compliance programs, but they give a team a concrete baseline to review and adjust rather than starting from a blank policy file.
A concrete scenario
Say an agent has access to a file-read tool and a network-request tool. Without a policy gate, a prompt injection in a document the agent reads could cause it to exfiltrate data by calling the network tool with file contents as a parameter.
With Vaikora Guard MCP in front of that agent, you write a policy rule: block any network-request tool call where the arguments contain content matching your data classification patterns. The proxy inspects the call arguments at runtime, matches against the rule, and blocks the call before it executes. The agent receives a refusal. The audit log records the blocked call with the matched rule and the full argument payload.
That is not a hypothetical attack type. Prompt injection via retrieved content is one of the most common risk patterns in agentic systems today.
What the open-source layer gives you
Publishing under MIT means you can read the inspection logic, verify how decisions are made, and confirm that no data leaves the proxy unexpectedly. For security teams evaluating third-party tooling, source availability changes the depth of review you can do.
The self-hosted architecture means call data, arguments, and audit logs stay inside your environment. The Vaikora Control Plane is the commercial product hosted by Data443, providing centralized policy management, dashboards, and SIEM integrations including Microsoft Sentinel, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. But the enforcement proxy itself runs where you run it.
The package supports 12 LLM providers with no SDK migration. If your agent stack already calls a hosted model API, the proxy sits in front of it without requiring changes to the agent code.
What to check next
If you want to evaluate the proxy before touching the control plane, install `vaikora-guard-mcp` from PyPI, point it at a test agent, and run a few tool calls through it. Review the audit log entries to confirm the decision structure matches what you expect.
Then write one rule that reflects an actual risk in your deployment, something specific to the tools your agents have access to. A policy gate with no rules is just a pass-through. The value comes from the first concrete constraint you enforce.