Most MCP servers execute tool calls and log what happened. vaikora-guard-mcp does the opposite: it intercepts the call, checks it against policy, and either allows, blocks, modifies, or holds it for approval before any execution occurs. The open-source package is available now on PyPI under the MIT license and is self-hostable.
What problem this solves
When an AI agent operates through the Model Context Protocol, it can call tools that touch real systems: file writes, database queries, API calls, external webhooks. The MCP specification defines how those calls are structured, but it does not define who decides whether a given call should run.
Without a policy gate, the agent and the LLM provider between them decide what executes. That means a prompt injection, a misconfigured agent, or a runaway loop can trigger destructive or data-exposing actions with no enforcement layer in the way.
vaikora-guard-mcp places itself inline as an MCP-compatible proxy. The agent talks to the proxy. The proxy evaluates the call. Only approved calls reach the actual tool server.
How it works mechanically
The proxy is written in Python and installs with a single command:
“`
pip install vaikora-guard-mcp
“`
Once running, it presents a standard MCP interface to the agent. When the agent issues a tool call, the proxy passes the call through six content inspection modules before forwarding it. Those modules check things like whether the call matches a blocked pattern, whether the arguments contain sensitive content, and whether the action falls within the policy scope you have configured.
The decision is deterministic: identical input produces identical output, every time. There is no probabilistic scoring that returns a different answer on retry. The call is allowed, blocked, modified to redact sensitive arguments, or held in a require-approval queue where a human can release or reject it.
Every decision is written to a SHA-256 hash-chained audit log. The chain is tamper-evident, so if a record is altered after the fact, the hash breaks. That property matters for compliance audits and for incident investigation, where you need an ordered, trustworthy record of what an agent did and when.
A concrete scenario
Suppose an agent is configured to help engineers query a production database. A developer asks it to "pull all user records for the last 90 days." Without a policy gate, the agent constructs a SQL query and executes it. If the table contains PII, that data is now in the LLM context, possibly logged by the model provider.
With vaikora-guard-mcp running inline, the proxy checks the tool call before it reaches the database connector. A policy rule that flags queries touching a table named `users` with no row limit can block the call and return a denial reason to the agent. The agent reports the block to the user. No data moves.
The same policy layer can require approval for calls that match a "high-risk" pattern, pausing execution until a human clears it through the control plane. That approval decision also appears in the audit log.
What the enterprise control plane adds
The open-source proxy handles local enforcement. The Vaikora Control Plane, hosted by Data443, adds centralized policy management across multiple agents, a dashboard for the approval queue, and the SIEM integrations.
Vaikora integrates natively with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub. Threats detected at the proxy layer push into those platforms as structured indicators, severity-mapped so they surface correctly in existing SOC workflows. That path requires no custom engineering on the SIEM side.
The control plane also ships compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. Rather than writing policy rules from scratch, you load a preset and adjust from there.
For teams that want the proxy without the control plane, the MIT license means you can run it entirely on your own infrastructure, modify it, and integrate it into your own tooling. The source is on PyPI as `vaikora-guard-mcp`.
Practical enforcement numbers
The proxy enforces at P50 8ms and P99 45ms per decision. At 10,000+ actions per second, it does not become the throughput ceiling for most agent deployments. Detection accuracy in testing is 99.9%. Those numbers come from Vaikora’s published test results, not extrapolation.
What to check next
If you are running MCP-connected agents in any environment that touches production data, pull the package, point it at your existing tool server, and watch what the audit log shows over 24 hours. The pattern of what your agents are actually calling, and what a policy layer would have blocked, is usually more informative than any theoretical risk assessment.
The PyPI page for `vaikora-guard-mcp` includes the configuration reference and a quickstart. Start there, then decide how much of the control plane you need.