VaikoraVaikora

Vaikora › Blog › Product Updates

Vaikora Guard MCP is now open source: here is what the policy check does on every tool call

Product Updates · Nov 20, 2026 · 4 min read

Tool calls are where AI agents cause real damage. A model that can read a file, write to a database, or call an external API can exfiltrate data, corrupt records, or trigger downstream effects that are difficult or impossible to reverse. The question is not whether to gate those calls. The question is where the gate lives and who controls it.

Vaikora Guard MCP puts that gate in a self-hosted proxy between the agent and every tool it calls. It is now published on PyPI as `vaikora-guard-mcp`, MIT-licensed, written in Python, and open to inspection, modification, and contribution.

What the proxy actually does

When an agent sends a tool call through the MCP protocol, the request hits the Vaikora Guard MCP proxy before it reaches the tool server. The proxy extracts the full context: the tool name, the input parameters, the agent identity, and the session state. It evaluates that context against the active policy set and returns one of four decisions: allow, block, modify, or require-approval.

The decision is deterministic. Identical input always produces identical output. There is no probabilistic scoring that might allow the same call to pass on Tuesday and block it on Wednesday. That consistency is what makes the audit log meaningful.

Every decision is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry is signed into the chain, so the ordered sequence of decisions cannot be altered without detection. If a compliance review or incident investigation asks what the agent was allowed to do at a specific time, the log answers that question with proof.

A concrete scenario

Consider an AI agent built to summarize customer support tickets. The agent has access to a file-reading tool and an email-sending tool. Without a policy gate, a prompt injection in a ticket could instruct the agent to read files outside the support directory and send their contents to an external address.

With Vaikora Guard MCP inline, a policy rule restricts the file-reading tool to paths under `/data/support/` and blocks the email-sending tool for any recipient domain not on an approved list. When the injected instruction runs, the file-read call to `/etc/secrets` is blocked before execution. The email call to the external address is blocked before execution. Both decisions appear in the audit log with the full input context, the rule that triggered, and the timestamp.

No alert fires after the fact. The action never executes.

Six content inspection modules

The proxy ships six content inspection modules: PII detection, secrets and credential scanning, prompt injection detection, URL reputation checks, topic and category filtering, and output content inspection. Each module runs at request time, before the tool call reaches its destination.

The enforcement latency is P50 8ms and P99 45ms. At that speed, the gate does not meaningfully affect agent throughput. The proxy can sustain more than 10,000 actions per second, which covers high-volume agentic pipelines without batching or queuing workarounds.

Compliance presets and the control plane

The open-source proxy enforces policy locally. The Vaikora Control Plane, the commercial product hosted by Data443, provides centralized policy management, team-level controls, compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001, and integrations with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.

The split is intentional. Security teams that want to inspect the enforcement logic, modify rules, or run entirely air-gapped can do so with the open-source proxy. Teams that need centralized governance across multiple agents, pre-built compliance configurations, and SIEM integration move to the control plane without changing the enforcement architecture.

Twelve LLM providers, no SDK migration

The proxy supports 12 LLM providers. Switching the enforcement layer on does not require migrating to a different SDK or rewriting agent code. The proxy sits at the MCP protocol layer, so it intercepts tool calls regardless of which model or framework the agent uses.

What to check next

If you are running agents that use MCP tool servers today, install `vaikora-guard-mcp` from PyPI and run it in logging mode first. Review the audit log entries for a representative sample of tool calls. Look at which tools are called most frequently, what parameters are passed, and whether any calls would have been blocked under a restrictive policy. That baseline is the starting point for writing policy rules that reflect what your agents should actually be allowed to do, not what the model decides to attempt.

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