VaikoraVaikora

Vaikora › Blog › Product Updates

What happens inside vaikora-guard-mcp when an AI agent tries to run a tool call

Product Updates · Nov 21, 2026 · 4 min read

The gap between "agent requests a tool call" and "tool call executes" is where policy enforcement has to live. If a check happens after execution, the action already ran. vaikora-guard-mcp is an open-source MCP-compatible proxy that sits in that gap and applies a deterministic allow, block, modify, or require-approval decision before the tool runs.

This post walks through what it actually does, how it is built, and what you get from running it yourself.

What the proxy does, mechanically

vaikora-guard-mcp intercepts the MCP message that would ordinarily pass directly from an AI agent to its tool server. Before that message reaches the tool, the proxy evaluates it against policy. The result is one of four outcomes:

Identical input always produces an identical decision. There is no probabilistic scoring involved in the enforcement step. If a policy says "block tool calls that include a Social Security Number in any parameter", that rule fires every time, not most of the time.

The package is published on PyPI as `vaikora-guard-mcp`, written in Python, and licensed MIT. You can self-host it without a commercial agreement.

Six content inspection modules

The proxy ships with six content inspection modules. Each module targets a specific category of risk in a tool call payload: credential and secret detection, PII identification, prompt injection patterns, URL and domain reputation checks, data exfiltration signals, and policy scope guardrails that restrict which tools an agent is permitted to call at all.

These run inline before the tool executes. If a tool call carries an AWS secret key in a parameter, the credential module catches it at inspection time, not in a post-execution log review.

Worked example: an AI agent is orchestrating a data pipeline and calls a file-write tool with a parameter that includes a bearer token it picked up earlier in the session. The credential inspection module identifies the token pattern, the proxy applies a block decision, and the write never happens. The agent receives a structured denial. The event is written to the audit log with a SHA-256 hash-chained entry, so the record is tamper-evident.

Audit log and the hash chain

Every decision, allow or block, is written to a SHA-256 hash-chained audit log. Each entry includes the input, the policy that matched, the outcome, and a hash that chains to the previous entry. This means the log cannot be silently edited. If someone removes an entry or alters a decision, the chain breaks and the tampering is detectable.

For compliance purposes, that matters. HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 all require demonstrable controls and auditability. An audit log that can be edited without detection does not satisfy that requirement. The hash chain structure closes that gap without any additional tooling.

Vaikora ships compliance presets for each of those frameworks, so the policy rules that match the control requirements are pre-configured rather than written from scratch.

Performance at the call layer

A proxy that adds meaningful latency to every tool call will get disabled. The enforcement overhead is P50 8ms, P99 45ms, and the gateway handles 10,000 or more actions per second. In practice, the inspection step adds less latency than most network round trips to the tool itself.

The proxy also supports 12 LLM providers without requiring an SDK migration. If your agent already calls a hosted model API, the adapter for that provider is included.

What the open-source layer gives you, and where the control plane fits

The MIT-licensed proxy gives you enforcement at the call layer, audit logging, and the six inspection modules. You can run it standalone against your own policy configuration files.

The Vaikora Control Plane (the commercial product, hosted by Data443) adds a centralized policy management interface, integrations with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub, and the ability to push AI agent threat signals into those platforms as native incidents or threat intelligence indicators. The open-source layer and the control plane share the same enforcement engine, so policies you define in either environment behave identically at runtime.

What to check next

If you want to evaluate the proxy against your own agent traffic, start with the PyPI package (`pip install vaikora-guard-mcp`) and run it in observe-only mode. Configure one inspection module at a time and review the audit log output before enabling block decisions. That sequence gives you a realistic picture of what the policy would have blocked in production before you commit to enforcement mode.

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

More from the Vaikora blog