Vaikora › Blog › Product Updates
Why open-sourcing the MCP proxy matters: policy enforcement at the tool call layer, before anything executes
Publishing source code is easy. Publishing the enforcement layer of an AI agent pipeline under MIT is a different kind of decision, because it means anyone can inspect exactly what the policy gate does and confirm it works the way the documentation claims.
That is what the release of vaikora-guard-mcp is actually about.
What the proxy does and where it sits
An AI agent using the Model Context Protocol issues tool calls: read this file, query this database, call this API. Those calls go out at machine speed, and nothing in the MCP specification itself enforces policy on them before execution.
vaikora-guard-mcp sits inline between the agent and every tool it calls. Before the tool call executes, the proxy evaluates it against your configured policy and returns one of four decisions: allow, block, modify, or require-approval. The tool call does not proceed until a decision is issued.
The enforcement logic is deterministic. Identical input produces identical output every time. There is no probabilistic scoring that lets a borderline call slip through on a good day. If the policy says block, it blocks.
What "open source under MIT" means in practice
The proxy is written in Python and published on PyPI as `vaikora-guard-mcp`. You can install it, read every line of enforcement logic, and run it in your own environment without depending on a hosted service.
For security teams, that matters for two reasons. First, you can audit the inspection logic yourself before trusting it in a production pipeline. Second, self-hosting means the proxy sees your agent traffic and nothing else does. No telemetry leaves your environment unless you configure it to.
A concrete setup looks like this: your agent is running against an LLM or a set of MCP tools, and you insert the proxy as the MCP server endpoint. Policy is defined in configuration. From that point, every tool call the agent issues hits the enforcement gate first. The agent never knows whether the call was allowed or blocked; it just receives the result or a denial, depending on the decision.
What you can enforce
The proxy ships six content inspection modules. These cover things like sensitive data patterns in tool call arguments, destination restrictions, and action types that require human approval before proceeding.
Compliance presets for SOC 2, HIPAA, GDPR, PCI DSS and ISO 27001 are included. A team subject to HIPAA can enable that preset and get a baseline policy that blocks tool calls carrying PHI patterns from reaching unapproved destinations, without writing the detection logic from scratch.
Every decision is written to a SHA-256 hash-chained audit log. The chain is tamper-evident: if any entry is modified after the fact, the chain breaks and the modification is detectable. For compliance audits or incident investigation, you have an ordered, verifiable record of every tool call decision the proxy made.
Where the enterprise control plane fits
The open-source proxy is self-hosted and self-managed. If your team needs centralized policy management across multiple agents, alert routing into your SIEM, or the native integrations with Microsoft Sentinel, CrowdStrike Falcon, and SentinelOne, that is the Vaikora Control Plane, the commercial product hosted by Data443.
The proxy and the control plane use the same enforcement logic and the same policy model. Teams commonly start with the open-source proxy to validate the approach in a development environment, then connect to the control plane when they move to production and need operational visibility across agents at scale.
Latency and throughput
A common concern with inline enforcement is that the proxy adds enough latency to break agent workflows. At P50 the enforcement decision takes 8ms. At P99 it takes 45ms. The proxy handles 10,000 or more actions per second. For most agent pipelines, those numbers mean the gate is not the bottleneck.
What to check next
If you are running agents over MCP today and have no policy gate on tool calls, install the proxy from PyPI (`pip install vaikora-guard-mcp`), configure a single policy rule, and run a test call through it. Watch what the audit log produces. That will tell you more about where your current agent pipeline has gaps than any threat model document will.
If you are evaluating whether to bring this into a larger deployment, review the enforcement logic in the repository directly. The source is there to be read.
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
Vaikora