Publishing a library as open source is easy. Making it actually enforce something before damage is done is harder. Vaikora Guard MCP, available on PyPI as `vaikora-guard-mcp` under an MIT license, sits inline between an AI agent and every tool call it wants to make, applying a policy decision before execution rather than logging what happened afterward.
This post explains the architecture, what the enforcement mechanism does, and how to reason about the audit trail it produces.
Why "before execution" matters
An AI agent operating through the Model Context Protocol can call file system tools, external APIs, database queries, or shell commands depending on what servers are connected. If a policy violation is detected after the call returns, the data has already moved, the query has already run, and the external request has already been sent. Post-execution detection is incident response, not prevention.
Vaikora Guard MCP operates as a proxy. The agent sends a tool call request. The proxy intercepts it, evaluates it against the configured policy, and returns one of four decisions: allow, block, modify, or require-approval. The underlying tool never receives the call unless that gate passes. That ordering is not a subtle distinction; it determines whether you have a governance control or a monitoring dashboard.
What the policy evaluation actually does
Policy is defined in a YAML file. Each rule names the action being controlled, the conditions under which it applies, and the enforcement decision. The evaluation is deterministic: identical input produces identical output every time. There is no probabilistic scoring that might allow the same call through on a different run.
The gateway ships six content inspection modules. These cover the content of the tool call itself, not just its name or category. A rule can inspect the payload of a database query, the destination of an outbound request, or the content of a file write, and block or modify based on what that content contains.
Consider a scenario: an agent is authorized to call a customer data API. A rule is configured to block any call where the payload contains a pattern matching raw PAN (payment card number) data. The agent constructs a call that happens to include card data it retrieved from an earlier step. The proxy inspects the payload, matches the rule, and blocks the call before the API receives it. The agent gets an error response. The card data never leaves the execution environment through that channel.
Audit chain and tamper evidence
Every decision the proxy makes is written to a SHA-256 hash-chained audit log. Each entry references the hash of the previous entry, so any modification to a past record breaks the chain and is detectable. This matters for compliance and investigation workflows where you need to prove what an agent did, what was blocked, and when.
Vaikora compliance presets ship for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. These are starting configurations, not finished policies, but they give a team something to audit against and modify rather than starting from a blank file.
Integration with security tooling
For teams running enterprise infrastructure, Vaikora integrates with Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. High-severity AI agent risk gets promoted into Azure Security Center and Microsoft Defender for Cloud as native incidents. Threat intelligence indicators are pushed into SentinelOne with severity scores mapped as: Critical=100, High=75, Medium=50, Low=25.
The enforcement throughput is 10,000-plus actions per second at P50 8ms and P99 45ms latency. For most agent deployments that means the policy gate does not become the bottleneck.
The proxy also supports 12 LLM providers without requiring SDK migration. It sits at the tool-call layer, so the model on the other end does not change how the enforcement works.
Self-hosting and the control plane
`vaikora-guard-mcp` is self-hosted. You run the proxy in your own environment, configure the YAML policy, and the enforcement happens locally. The MIT license means you can inspect the code, modify it, and integrate it into your own tooling without restriction.
The commercial product is the Vaikora Control Plane, hosted by Data443. It adds centralized policy management, the tamper-evident audit log at scale, the SIEM integrations, and the compliance preset library. Teams that need to manage policy across many agents or many environments, or that need the audit log to meet a regulatory requirement, will generally need the control plane alongside the open-source proxy.
What to check next
If you are running AI agents with MCP tool access and you do not have a policy gate in front of those tool calls, the starting point is installing `vaikora-guard-mcp` from PyPI and reviewing the default policy file against the tools your agents can reach. Focus first on any tool that writes data, calls an external API, or executes commands. Those are the calls where post-execution detection is too late to matter.