Most AI agent frameworks give you logging after the fact. What they do not give you is a way to block an action before it runs, based on a deterministic policy, with a tamper-evident record of every decision. That gap is what vaikora-guard-mcp is built to close.
The package is MIT-licensed, written in Python, and published on PyPI. You can install it today with `pip install vaikora-guard-mcp`. The source is public. There is no vendor lock-in at the gateway layer.
What the proxy actually does
vaikora-guard-mcp sits inline between your AI agent and the tools it calls through the Model Context Protocol. Every tool call passes through a policy check before execution. The gateway returns one of four decisions: allow, block, modify, or require-approval. The action either proceeds, is rewritten, is held for a human reviewer, or is stopped. Nothing executes before that decision is made.
This matters because most guardrail approaches work after the fact. They detect bad outputs, log anomalies, or alert a SOC analyst. By the time you get the alert, the tool call has already run. The database write happened. The API request went out. With an inline proxy, enforcement happens at the only moment where it can still change the outcome.
The gateway ships six content inspection modules. They cover things like sensitive data patterns, prompt injection attempts, and policy violations against the action payload. You configure what each module checks and what threshold triggers which decision type.
A concrete example: file exfiltration guardrail
Suppose you have an agent with access to a filesystem tool and a network tool. Without a policy gate, nothing stops the agent from reading a file containing PII and then passing its contents to an outbound HTTP call. The two tool calls are individually permitted. The combination is the problem.
With vaikora-guard-mcp, you can write a policy that inspects the payload of any tool call with a network destination. If the content inspection module detects a PII pattern in the outgoing payload, the gateway returns a block decision before the HTTP call executes. The agent receives a rejection, not a successful response. The data does not leave the environment.
That policy is configured in YAML. It is version-controllable, reviewable in a pull request, and auditable in the decision log.
Audit chain and what it means for compliance
Every decision the gateway makes is written to a SHA-256 hash-chained audit log. Each entry references the previous one, so any modification to a historical record breaks the chain and becomes detectable. This is not just logging. It is an ordered, tamper-evident record of every policy decision, with the input, the policy that matched, and the outcome.
For teams operating under HIPAA, PCI-DSS, GDPR, SOC 2, or ISO 27001, this matters. Compliance presets for each of those frameworks ship with the platform. When an auditor asks what controls exist on AI agent actions, the answer is a policy file and a verifiable decision log, not a monitoring dashboard and a promise.
Open-source gateway, enterprise control plane
The gateway (vaikora-guard-mcp) handles enforcement locally, in your own environment. The commercial product is the Vaikora Control Plane, hosted by Data443, which adds centralized policy management, a UI for approval workflows, SIEM integrations, and reporting across multiple agent deployments.
If you run the open-source proxy alone, you get enforcement and local audit logs. If you connect it to the control plane, you get threat indicators pushed into SentinelOne (severity-mapped at Critical=100, High=75, Medium=50, Low=25), incidents promoted into Microsoft Defender for Cloud, and telemetry flowing into Microsoft Sentinel, CrowdStrike Falcon, and AWS Security Hub.
The practical split: start with the open-source package to validate that inline enforcement fits your agent architecture. Add the control plane when you need centralized governance across more than one agent or team.
Latency and throughput at the enforcement layer
A common objection to any inline proxy is performance. The gateway operates at P50 8ms and P99 45ms per enforcement decision, with throughput above 10,000 actions per second. For most agent workflows, that overhead is not meaningful. For high-volume pipelines, the architecture supports horizontal scaling.
What to do next
Pull the package from PyPI and read the README before you configure anything. The configuration format is YAML and the inspection modules are individually togglable, so you can start with a single module in audit-only mode to see what your agents are actually doing before you turn on blocking.
If you are running agents against MCP-compatible tools in a regulated environment, check whether your current logging setup can prove that a specific tool call was evaluated against a specific policy version at a specific time. If it cannot, that is the gap vaikora-guard-mcp is designed to fill.