Vaikora › Blog › Product Updates

We put the Vaikora policy engine on PyPI so any MCP agent can enforce before it executes

Product Updates · Nov 6, 2026 · 3 min read

An AI agent that can call tools without a policy gate will eventually call the wrong one. The question is whether you find out from a log review or from blocking it in real time.

That is why we published vaikora-guard-mcp as an MIT-licensed Python package on PyPI. Any team running an MCP-compatible agent can install it, drop it inline between the agent and its tools, and have deterministic policy enforcement running in minutes. No enterprise contract, no SDK migration, no custom integration work.

What the proxy actually does

vaikora-guard-mcp sits between the MCP client and the tool servers the agent would otherwise reach directly. When a tool call arrives, the proxy evaluates it against your policy before the call executes. The result is one of four decisions: allow, block, modify, or require-approval. The tool either runs or it does not. There is no race condition between detection and action.

The enforcement decision is deterministic. Identical input produces identical output every time. That matters for auditability: when an incident reviewer asks why a specific tool call was blocked at a specific time, the answer is in the policy, not in probabilistic scoring.

Every decision is written to a SHA-256 hash-chained audit log. The chain is tamper-evident, so the record you produce for a compliance review is the same record that existed at the time of the event.

Installing it

```
pip install vaikora-guard-mcp
```

Point your MCP client at the proxy instead of your tool servers. The proxy handles routing to the actual servers after policy evaluation. Your agent code does not change.

A minimal configuration file describes which tool namespaces are in scope and what the default action is for unmatched calls. From there you add specific rules: block calls to any file-write tool that targets paths outside a defined prefix, require approval for any tool that initiates an outbound network connection, allow everything else.

A concrete scenario

Consider an agent that has access to three tools: a document reader, a code executor, and an HTTP client. Without a policy gate, the agent can chain those tools in ways that were never intended, reading a file, executing code that processes it, and sending results to an external endpoint.

With vaikora-guard-mcp inline, you write one rule: block any tool call where the tool name matches the HTTP client and the destination is not on an approved list. The proxy evaluates that rule at P50 8ms and P99 45ms. The agent never sees a response from the blocked tool. The attempt and the block are both in the audit log.

That log entry includes the tool name, the parameters, the policy rule that matched, and the hash linking it to the entry before and after. If a compliance auditor asks whether any outbound calls were made during a specific session, the answer is provable, not inferred.

Connecting to the control plane

The open-source proxy is self-hosted. If your team needs centralized policy management across multiple agents, shared approval workflows, or the compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001, the Vaikora Control Plane connects to the same proxy. You switch from local policy files to centrally managed policy without changing the agent or the proxy deployment.

The proxy also integrates natively with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. High-severity enforcement events appear in those platforms as incidents or threat intelligence indicators, with no custom engineering on your end.

What to check next

If you are running any MCP-compatible agent today, the first thing to verify is what happens when that agent calls a tool that it should not be allowed to call. If the answer is "the log captures it after the fact," you have a detection problem, not an enforcement problem.

Install vaikora-guard-mcp, run it in audit-only mode against your existing agent traffic, and review what the policy log surfaces. That gives you a factual baseline before you turn enforcement on.

The package is on PyPI and the source is on GitHub under the MIT licence. Start there.

More from the Vaikora blog