Shipping an AI agent without a policy gate means every tool call runs on trust. If the agent decides to write a file, query a database, or call an external API, nothing stops it except the logic you wrote into the prompt. That gap is why we open-sourced vaikora-guard-mcp under the MIT licence.
This post covers what the package actually does, how to get it running, and what policy enforcement looks like in practice.
What the Package Is
vaikora-guard-mcp is a self-hostable MCP-compatible proxy written in Python. It sits inline between an MCP client (your AI agent or orchestration layer) and the tools that client can call. Before any tool call executes, the proxy evaluates it against a configured policy and returns one of four decisions: allow, block, modify, or require-approval.
It is published on PyPI as `vaikora-guard-mcp` and the source is MIT licensed. You can inspect every line, fork it, and run it on your own infrastructure. There is no telemetry phoning home unless you wire it to the enterprise control plane.
How Policy Enforcement Works
The proxy intercepts the tool call at the protocol layer before execution. It reads the tool name, the arguments, and any context the agent passes. It then evaluates that payload against your policy rules deterministically: identical input always produces identical output. There is no probabilistic scoring on the enforcement path.
A concrete example: an agent calls `write_file` with a path argument of `/etc/hosts`. Your policy blocks any `write_file` call targeting system directories. The proxy returns a block decision with a reason string. The tool never runs. The agent receives the rejection and can respond accordingly, log it, or escalate.
Decisions are written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so you have a tamper-evident record of every allow and every block. That matters when a compliance auditor asks what your agent did at 2:47 AM last Tuesday.
What You Get Out of the Box
The gateway ships six content inspection modules. These cover common policy surfaces: sensitive data patterns in arguments, tool call rate controls, argument schema validation, blocklists for specific tool names or endpoints, approval workflows for high-risk actions, and context-aware rules that consider the calling agent’s identity.
Compliance presets for SOC 2, HIPAA, GDPR, PCI-DSS, and ISO 27001 are included. If you are operating under one of those frameworks and need to demonstrate that your AI agent layer has controls, the preset gives you a documented starting point rather than a blank configuration file.
Enforcement latency runs at P50 8ms and P99 45ms, so the gate does not add meaningful delay to agent workflows operating at human interaction speed.
Running It Locally
“`
pip install vaikora-guard-mcp
“`
From there, point your MCP client at the proxy endpoint instead of directly at the tool server. The README covers the configuration schema for defining policy rules and wiring up the audit log output. You can start with the default-deny preset and incrementally add allow rules for the specific tools your agent actually needs. This is the opposite of default-allow, which is what most deployments start with and later regret.
Where the Open-Source Package Ends
vaikora-guard-mcp handles enforcement at the single-proxy level. If you need centralized policy management across multiple agents, a UI for the approval workflow, integrations with Microsoft Sentinel, CrowdStrike Falcon, or SentinelOne, or the hosted control plane that ties decisions across environments, that is the enterprise layer Data443 operates as a SaaS subscription. It is also available on AWS Marketplace with EDP draw-down support.
The open-source package is intentionally complete for teams that want self-contained enforcement. The enterprise layer is for teams that need governance across a fleet of agents.
What to Check Next
If you are evaluating whether the proxy fits your environment, the first practical step is to run it against a tool call you already make in development and check whether the audit log output matches what you expect. The hash chain is the part worth verifying early: open two entries and confirm the previous hash in the later entry matches the hash of the earlier one. If that holds, you have evidence the log has not been modified. If you are integrating with an existing SIEM, the Vaikora documentation covers the Sentinel and security platform connector configuration.