Vaikora › Blog › Product Updates
Vaikora Guard MCP is now open source: here is what that means for AI agent policy enforcement
Policy checks for AI agents have mostly been implemented after the fact: log the tool call, inspect the output, alert if something looked wrong. By then the action has already run. Vaikora Guard MCP flips that order. It sits inline between an agent and the tools it calls, and applies a policy decision before execution. That code is now publicly available on PyPI under the MIT licence.
What Guard MCP is and where it sits
Vaikora Guard MCP is a self-hostable MCP-compatible proxy written in Python. Install it as `vaikora-guard-mcp` from PyPI. Once in place, every tool call an agent attempts passes through it. Guard MCP evaluates that call against configured policy, then returns one of four deterministic decisions: allow, block, modify, or require-approval. The action does not proceed until a decision is made.
That placement matters. The proxy is not watching a log stream. It is in the request path. An agent that tries to exfiltrate data, call an unauthorized API, or write to a restricted resource is stopped before the tool executes, not after.
What "deterministic" means in practice
Deterministic policy enforcement means the same input always produces the same decision. There is no probabilistic scoring that might allow an action today and block it tomorrow with no configuration change. If a policy rule says "block any tool call that includes a Social Security Number pattern," every call containing that pattern is blocked. No variance.
Guard MCP ships six content inspection modules that run against each tool call. These cover things like sensitive data patterns, prompt injection attempts, and policy-defined topic restrictions. The modules run in sequence before the action reaches the LLM or the external tool.
A concrete example: an agent working inside a customer support workflow tries to call a file-retrieval tool and passes a parameter that contains what looks like a credit card number. Guard MCP's content inspection catches the pattern, the policy rule evaluates to block, and the tool call is rejected. The agent receives a policy rejection rather than a result. The whole exchange is written to a SHA-256 hash-chained audit log before the response is returned.
The audit record
Every decision Guard MCP makes is written into a tamper-evident, SHA-256 hash-chained audit log. Each entry is signed into the chain, so any attempt to delete or alter a past decision breaks the chain integrity. For compliance investigations, this gives you an ordered, verifiable record of what every agent attempted, what policy evaluated, and what the outcome was.
This matters for SOC 2, HIPAA, GDPR, PCI DSS, and ISO 27001 requirements where you need to demonstrate that controls were applied and decisions were made consistently. Guard MCP ships compliance presets for all five. The preset loads a baseline policy configuration appropriate for that framework; you adjust from there.
Performance numbers worth knowing
An inline enforcement layer is only usable if it does not add unacceptable latency. Guard MCP enforces policy at P50 of 8ms and P99 of 45ms, with the gateway holding sub-500ms at P95. At throughput, it handles over 10,000 actions per second. For most agent workflows, that overhead is invisible.
What open source means operationally
MIT licence means you can read the enforcement logic, audit it, modify it, and run it in your own environment with no dependency on Data443's infrastructure. The gateway itself is self-hosted. The commercial product is the Vaikora Control Plane, which adds centralized policy management, the enterprise integrations (Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, AWS Security Hub), and the managed control layer.
If you want to evaluate the enforcement behavior in isolation, you can do that entirely within your own infrastructure. The PyPI package is the starting point.
Guard MCP supports 12 LLM providers with no SDK migration required. If you are already calling hosted model APIs directly, the proxy sits in front of those calls without requiring you to rewrite your agent code.
What to check before you deploy
Before you install Guard MCP in an agent workflow, map your tool call surface: what tools the agent can call, what parameters each tool accepts, and what data types could appear in those parameters. That mapping becomes the basis for your first policy rules.
Then decide which policy decisions need a human in the loop. Guard MCP's require-approval action holds the call and waits for explicit authorization before proceeding. For high-risk tool categories, that gate is worth configuring from the start.
The repository and PyPI package are the right place to start. Review the content inspection modules, run the enforcement logic against sample tool calls from your own agents, and check how the audit chain output maps to your existing log pipeline before you go further.
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