Vaikora › Blog › Product Updates
Every tool call your AI agent makes can be checked before it runs. Here is the open-source proxy that does it.
When an AI agent calls a tool, that call executes immediately unless something sits between the agent and the tool and makes a decision. Without that gate, every file write, API call, database query, and shell command runs on trust alone.
Vaikora Guard MCP is the open-source component that closes that gap. It is an MIT-licensed, self-hosted, MCP-compatible proxy written in Python and published on PyPI as `vaikora-guard-mcp`. It sits inline between an agent and any tool it invokes, and it enforces a policy decision before execution. Not after. Not in a log reviewed hours later.
What the proxy actually does
Every tool call passes through Vaikora Guard MCP as an inspection request. The proxy applies your policy file and returns one of four decisions: allow, block, modify, or require-approval. The action proceeds only if the decision permits it. Identical input always produces an identical decision, so the enforcement is deterministic, not probabilistic.
The proxy enforces at P50 8ms and P99 45ms, so the latency added to each tool call is well within what agents can absorb without meaningful performance change. At scale, the enforcement path handles 10,000 or more actions per second.
Every decision is written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so the log is tamper-evident and ordered. If a tool call was blocked on Tuesday at 14:23, that record cannot be modified or removed without breaking the chain. That is the property that matters for compliance and incident investigation.
MCP compatibility and what it means in practice
MCP (Model Context Protocol) is the emerging standard for how agents discover and call tools. Supporting MCP means the proxy works with any agent that speaks the protocol, without changes to the agent code or the tool implementations. You do not change your SDK. You point your agent at the proxy endpoint instead of the tool directly.
A concrete example: an agent uses a file-system tool to write output. Without a gate, the write executes immediately. With Vaikora Guard MCP deployed, the write request hits the proxy first. The proxy checks the destination path, the content, and any configured guardrails, then issues its decision. If the policy says block writes to `/etc` or any path matching a sensitive data pattern, the write never happens.
The proxy ships with six content inspection modules covering data classification, prompt injection patterns, PII detection, and tool-call structure validation. Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 are included, so you are not writing policy from scratch for regulated workloads.
Self-hosting and the control plane split
The proxy is self-hosted. Your tool call traffic does not leave your environment. That is the point of the architecture: enforcement happens inside your perimeter, against your policy, with your audit log stored where you control it.
The commercial side, the Vaikora Control Plane hosted by Data443, adds policy management across multiple agents, centralized audit search, SIEM integrations (Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, AWS Security Hub), and the enterprise approval workflow. The open-source proxy works independently if you want to manage policy through the YAML config directly.
This split is intentional. Teams that need to validate the enforcement logic, audit it, or modify it can do so against the open-source code. The MIT license places no restrictions on how you deploy or integrate it.
Detection accuracy and what it is based on
Vaikora Guard MCP reports 99.9% detection accuracy in testing. That figure applies to the content inspection modules identifying the patterns they are configured to catch. The practical implication is that a policy rule set to catch tool calls that touch credit card numbers or PHI will catch them at that rate before execution, not after a post-hoc log review.
What to check next
If you are running AI agents in any environment where tool calls touch sensitive data, external APIs, or systems with compliance scope, the first thing to verify is whether anything sits between the agent and those tools right now.
If the answer is nothing, start with the open-source proxy. It is on PyPI, the configuration is YAML, and the first policy checkpoint can be running in an afternoon. Check the vaikora-guard-mcp repository for the full setup documentation, the inspection module reference, and the compliance preset files.
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
Vaikora