Vaikora › Blog › Product Updates

What vaikora-guard-mcp actually enforces, and why we published it under MIT

Product Updates · Nov 13, 2026 · 4 min read

A tool call is the moment an AI agent stops reasoning and starts acting. File written, API called, query executed. Policy that only watches after that point is not enforcement, it is a log. vaikora-guard-mcp intercepts the call before it runs and applies a deterministic allow, block, modify, or require-approval decision. We published the source under MIT so anyone can read exactly what happens at that gate.

What the proxy does, mechanically

vaikora-guard-mcp sits inline between an MCP client (your agent) and whatever tools the agent can call. Every tool call passes through six content inspection modules before it reaches the tool. The modules check things like data classification signals in the payload, whether the target matches a policy rule, and whether the action type is permitted for the current context.

The decision is deterministic: identical input always produces identical output. There is no probabilistic scoring that lets the same call through one minute and blocks it the next. This matters for auditability. If you need to prove to an auditor that a specific action was blocked on a specific date, the record is there, signed into a SHA-256 hash-chained audit log, and the decision cannot be altered after the fact.

Latency at P50 is 8ms. At P99 it is 45ms. For the vast majority of tool calls, the enforcement check is invisible in wall-clock terms.

Why open source

The enforcement path in an AI agent is exactly the kind of code that should be inspectable. If your security team cannot read the source of the thing that is supposed to stop a data exfiltration call, they are trusting a black box in a critical position. MIT licensing means you can read it, modify it, run it in an air-gapped environment, and contribute fixes back.

The package is published on PyPI as `vaikora-guard-mcp`. The gateway is written in Python. You can install it, point it at your MCP server, write a YAML policy file, and have a working policy gate running locally.

What a policy file looks like in practice

Say you have an agent that can call a file-write tool and a web-search tool. A minimal policy might look like:

• Block any file-write action where the payload matches a PII classification signal (SSN, credit card pattern, health record field)
• Require human approval for any file-write that targets a path outside a defined working directory
• Allow web-search unconditionally
• Block web-search if the target domain is on a threat intelligence blocklist

Each rule is evaluated in order. The first matching rule wins. The decision (allow, block, modify, require-approval) is written to the audit chain along with the full action payload, the matched rule, and a timestamp. You get an ordered, tamper-evident record of every decision the gate made.

Compliance presets

For teams that need to map policy to a framework quickly, vaikora-guard-mcp ships presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. A preset loads a baseline rule set appropriate for that framework. You can extend or override any rule. The preset is a starting point, not a ceiling.

Where the enterprise control plane fits

The open-source proxy handles enforcement at the agent level. The Vaikora Control Plane, hosted by Data443, adds centralized policy management across multiple agents, a unified audit view, SIEM integration (Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, AWS Security Hub), and threat-level mapping for high-severity decisions. Teams running more than one agent workflow typically want the control plane so policy changes do not require touching each deployment individually.

The control plane is optional. The proxy enforces policy with or without it.

What to check next

• Clone or install `vaikora-guard-mcp` from PyPI and run it against a local MCP server with a test tool. Confirm that a blocked call does not reach the tool and that the audit entry appears with the correct rule reference.
• Review the default rule evaluation order in the source. If your environment has overlapping rules, the ordering behavior is the thing most likely to surprise you.
• If you are running agents under service accounts with broad IAM or RBAC permissions, pair the policy gate with a review of what those accounts can actually reach. The gate blocks calls that match policy, but it cannot block calls that your policy does not cover.

The source is at vaikora-guard-mcp on PyPI.

More from the Vaikora blog