Vaikora › Blog › Product Updates

vaikora-guard-mcp: the MIT-licensed MCP proxy that enforces policy before a tool call runs

Product Updates · Nov 8, 2026 · 3 min read

An AI agent that can call tools is an agent that can cause side effects. The question is not whether to gate those calls, it is where the gate sits and whether it is auditable when something goes wrong.

vaikora-guard-mcp is a small, self-hostable MCP-compatible proxy written in Python and published on PyPI under the MIT licence. It sits inline between an AI agent and the tools it calls. Before any tool call executes, the proxy evaluates it against policy and returns one of four decisions: allow, block, modify, or require-approval. The tool either runs or it does not. There is no post-execution cleanup.

What the proxy actually intercepts

When an MCP client sends a tool call, vaikora-guard-mcp receives the request first. It passes the full call, including tool name, arguments, and any attached context, through a set of content inspection modules before forwarding it downstream. If the call matches a block rule, the agent gets a structured refusal. If it matches a modify rule, the proxy rewrites the call before execution. If it matches a require-approval rule, the call pauses until a human responds.

This happens at P50 8ms and P99 45ms, so the latency added to any single tool call is not meaningful in practice. The proxy handles over 10,000 actions per second, which covers any realistic agentic workload.

Installing and connecting it

```
pip install vaikora-guard-mcp
```

Point your MCP client at the proxy instead of the tool server directly. No SDK migration is required. The proxy adapts to twelve LLM providers, so the agent layer does not need to change. The gateway ships six content inspection modules covering topics including sensitive data patterns, prompt injection patterns, and tool scope restrictions.

A concrete scenario: an agent authorized to query a database is also capable of issuing writes if the tool schema allows it. A policy rule scoped to that tool can block any call that includes a mutation keyword in the arguments, regardless of what the agent was instructed to do. The block fires before execution, not after the write lands.

The audit record

Every decision the proxy makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry is signed into the chain, so the log produces an ordered, verifiable record of what was allowed, what was blocked, and when. If a compliance audit or incident investigation asks what a specific agent did during a specific window, the answer is in the log and the log cannot be quietly edited.

This matters for teams operating under HIPAA, PCI-DSS, GDPR, SOC 2, or ISO 27001. Vaikora ships compliance presets for all five, which map common regulatory requirements to concrete policy rules rather than leaving that translation to the team implementing it.

What the open-source boundary is

vaikora-guard-mcp is the enforcement proxy. It is MIT-licensed and self-hosted. The Vaikora Control Plane, the commercial product hosted by Data443, adds centralized policy management, multi-agent governance, and the integrations that push AI agent threat signals into Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub. Teams that want a single enforcement point with no external dependency can run the proxy alone. Teams that need fleet-wide policy control or SIEM-level visibility connect it to the control plane.

The proxy also adds a runtime gate on top of whatever IAM controls already exist. Azure RBAC, AWS IAM, and Snowflake roles still apply. The proxy enforces policy before a call reaches the point where those controls would be evaluated, which means a policy violation never touches the underlying permission boundary at all.

What to check next

Pull the package, run it against a test tool server, and read what the inspection modules log on a few representative tool calls from your existing agents. The audit log entries will show you which calls would have been blocked or modified under a default policy. That output is a reasonable starting point for writing policy rules that reflect what your agents should actually be allowed to do.

The source is on PyPI as vaikora-guard-mcp. The repo includes configuration examples and documentation for each inspection module.

More from the Vaikora blog