Vaikora › Blog › Product Updates

How to add a policy enforcement layer to your MCP server in under ten minutes

Product Updates · Oct 22, 2026 · 3 min read

The default behavior of any MCP server is to pass tool calls straight to execution. No check, no record, no way to prove after the fact what ran and what did not. For most production environments that is not acceptable, but wiring up a custom enforcement layer from scratch takes time most AI platform teams do not have.

Vaikora Guard MCP solves that with a self-hostable proxy you can place inline between your agent and the tools it calls, published on PyPI as `vaikora-guard-mcp` under the MIT licence.

What the proxy actually does

`vaikora-guard-mcp` sits between the agent and every tool call. Before execution, it checks the call against a policy and returns one of four deterministic decisions: allow, block, modify, or require-approval. The call does not proceed until a decision is made. This is enforcement before execution, not detection after the fact.

The proxy ships six content inspection modules. These cover the payload of the call itself, not just its metadata, so a call that looks structurally valid but carries exfiltration-pattern content is caught at the same point as one with an outright policy violation in the tool name.

Every decision is written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so the record is tamper-evident. You can produce that log for a compliance audit and demonstrate the exact sequence of decisions, in order, without gaps.

Setting it up

Install from PyPI:

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

Point your agent's MCP client at the proxy endpoint instead of directly at your tool server. The proxy forwards the call to the real tool server only if the policy decision is allow or modify. No SDK migration is required. Vaikora supports 12 LLM providers through the same interface.

A minimal policy file defines which tool names are permitted, which are blocked outright, which require a human approval step, and what content patterns in the payload trigger a block or modification. The policy is evaluated deterministically: identical input always produces the same decision. That predictability is what makes the audit log meaningful.

A concrete scenario

An AI agent is wired to a file-management tool server. The agent, responding to a user prompt, constructs a tool call that would write a file to an external path outside the permitted directory. Without a policy layer, that call runs. With `vaikora-guard-mcp` in front of it, the call hits the proxy, the path check fails against the policy, the decision is block, the call never reaches the tool server, and the decision is appended to the audit chain with a timestamp and the full call payload.

The developer sees the blocked call in the audit log. The compliance team can pull that log and show an auditor every decision the system made over a given period, in sequence, with tamper-evident integrity.

What the enterprise control plane adds

The open-source proxy handles enforcement. The Vaikora Control Plane is the commercial layer that adds centralized policy management across multiple agents, real-time dashboards, compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001, and integrations with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub.

For teams running a single agent in a controlled environment, the open-source proxy alone may be sufficient. For teams running multiple agents in production, the control plane is where policy becomes consistent across the fleet rather than maintained per-agent.

Enforcement latency at P50 is 8ms, and the proxy handles more than 10,000 actions per second. Adding the proxy does not meaningfully change the performance profile of a typical agent workflow.

What to check next

• Clone the repo or `pip install vaikora-guard-mcp` and read the policy schema. It takes less time than writing a custom middleware.
• Decide which tool calls in your current setup have no enforcement at all. Those are the calls that need a policy entry first.
• If you need tamper-evident records for a compliance program, confirm that your policy file covers every tool registered in your MCP server, not just the ones you expect the agent to call.

The source is on PyPI and the licence is MIT, so you can inspect everything before running anything.

More from the Vaikora blog