VaikoraVaikora

Vaikora › Blog › Product Updates

One proxy, one policy file, zero SDK changes: the vaikora-guard-mcp open-source release explained

Product Updates · Nov 17, 2026 · 4 min read

The gap that matters in an AI agent pipeline is not at the model. It is between the model's decision to call a tool and the moment that tool actually runs. If nothing sits in that gap, every policy rule you write lives only in documentation.

vaikora-guard-mcp is an MIT-licensed Python proxy, published on PyPI, that closes that gap by sitting inline on the Model Context Protocol path. Each time an agent calls a tool, the proxy intercepts the call, evaluates it against your policy, and either allows it, blocks it, modifies the payload, or holds it for human approval before anything executes.

What the proxy actually does

The proxy acts as a compliant MCP server. Your agent connects to it exactly as it would connect to any MCP-compatible endpoint. The proxy forwards approved calls downstream and returns results normally. From the agent's perspective, the proxy is invisible on the happy path.

On the enforcement path, the story is different. A tool call that matches a block rule never reaches the downstream system. The agent receives a structured refusal. A call that matches a require-approval rule is held in queue until a human responds. The agent waits, or times out, depending on how you configure it.

Every decision, allow or block or modify, is written to a SHA-256 hash-chained audit log. The chain is tamper-evident, so you can produce an ordered record of every tool call and its outcome for a compliance audit without reconstructing it from scattered logs.

Why deterministic enforcement matters

The proxy enforces policy deterministically. Identical input always produces identical output. There is no model scoring a risk level and returning a probability. The rule either matches or it does not.

A concrete example: you write a rule that blocks any tool call where the arguments contain a string matching your internal account number format. Every call with that pattern is blocked, every time, regardless of which agent sent it, which model generated the call, or what time of day it is. That predictability is what makes the control auditable. You can point a compliance reviewer at the rule and at the log entry and show them exactly why a call was blocked.

Installing and configuring it

The package is on PyPI as `vaikora-guard-mcp`. Install it, write a policy file in YAML that lists your allow, block, modify, and require-approval rules, and point your agent at the proxy's MCP endpoint instead of the tool's native endpoint. No SDK migration is required.

The proxy ships six content inspection modules: you can inspect argument names, argument values, tool names, call patterns, and payload structure. Rules compose, so you can build layered conditions without writing custom code.

For teams running the enterprise control plane, policy syncs centrally and the audit log feeds into the control plane's dashboard. For teams running fully self-hosted, the audit log is local.

Where the open-source boundary is

vaikora-guard-mcp, the proxy itself, is open source under the MIT licence. The Vaikora Control Plane, which provides centralized policy management, the enterprise dashboard, compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001, and native integrations with Microsoft Sentinel, CrowdStrike Falcon, and SentinelOne, is the commercial product hosted by Data443.

The split is intentional. A security team that wants to inspect and modify the enforcement logic can do that directly in the open-source code. An organization that needs centralized governance across multiple agents, auditability at scale, and pre-built compliance configurations can add the control plane.

What to check before you deploy

Before you route production agent traffic through any inline proxy, confirm three things:

The source is on PyPI as `vaikora-guard-mcp`. Read the policy file schema before writing your first rule. That is the fastest path to understanding what the proxy will and will not catch by default.

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

More from the Vaikora blog