Every time an AI agent calls a tool, something has to decide whether that call is allowed. Without a control at that point, the agent runs whatever the LLM produces. Vaikora Guard MCP is a self-hosted, MIT-licensed proxy that places a deterministic policy check at exactly that point, before execution, and it is now published on PyPI as `vaikora-guard-mcp`.
What the Proxy Actually Does
Vaikora Guard MCP sits inline between an AI agent and the tools it calls via the Model Context Protocol. When an agent produces a tool call, the proxy intercepts the request, evaluates it against your configured policy, and returns one of four decisions: allow, block, modify, or require-approval. The action does not proceed until a decision is issued.
That ordering matters. Detection after execution tells you something went wrong. Enforcement before execution prevents the action from completing in the first place. For agents that can write to databases, call external APIs, send email, or query filesystems, the difference is not subtle.
The proxy is written in Python, so it fits into most existing deployment patterns without additional runtime dependencies. Because it is MCP-compatible, it works with any agent that already speaks MCP, with no SDK migration required.
Deterministic Policy, Not Probabilistic Filtering
One design decision worth understanding: Vaikora applies deterministic policy enforcement. Identical input always produces an identical decision. That is different from using a second LLM to judge whether a tool call looks acceptable, where the same input might produce different outputs on different runs.
Determinism matters for two reasons. First, you can reason about what your policy actually covers, because the rules are explicit and testable. Second, you get a stable audit record. Every decision the proxy makes is written to a SHA-256 hash-chained, tamper-evident audit log. If a regulator or internal audit team asks what an agent did and why a particular call was blocked or approved, the log answers that question with a verifiable record rather than a reconstruction.
Six Content Inspection Modules
Vaikora Guard MCP ships with six content inspection modules covering the categories of content most likely to appear in agent tool calls that need policy guardrails: prompt injection attempts, sensitive data patterns, restricted topic areas, output format constraints, and others in the set. Each module evaluates the payload independently, so you can enable only what your use case requires and leave the rest off.
A worked example: an agent running in a healthcare workflow calls a tool that queries a patient record store. Before that call executes, the proxy checks whether the payload contains data patterns that would violate a HIPAA guardrail. If it does, the call is blocked and the decision is logged. If it does not, the call proceeds. The agent gets a response indicating what happened either way.
The enforcement latency at P50 is 8ms and at P99 is 45ms, so the check adds no meaningful delay to workflows operating at human scale, and the proxy sustains more than 10,000 actions per second for high-volume pipelines.
Compliance Presets
Vaikora ships preconfigured compliance presets for SOC 2, HIPAA, GDPR, PCI DSS, and ISO 27001. For teams that need a defensible policy baseline quickly, those presets reduce configuration time significantly. They are starting points, not locked configurations; you can extend them or write custom rules on top.
The Enterprise Control Plane
The open-source proxy handles enforcement and logging at the agent level. The commercial product is the Vaikora Control Plane, hosted by Data443, which adds centralized policy management, aggregate visibility across multiple agents and deployments, and integration with security tooling including Microsoft Sentinel, Microsoft Defender, SentinelOne, CrowdStrike Falcon, and AWS Security Hub.
For teams that want to start with the open-source proxy and evaluate whether the control plane adds value for their environment, that path is straightforward: deploy `vaikora-guard-mcp`, run it against a real agent workflow, and inspect the audit log to see what the enforcement record looks like before deciding whether centralized management is necessary.
What to Check Next
If you have an agent in production or staging that calls external tools, the first question to answer is what policy governs those calls today. If the answer is "the LLM decides," that is the gap. Install `vaikora-guard-mcp` from PyPI, point it at one agent, and read the audit log after a day of traffic. The log will show every tool call, the decision applied, and the reason. That output is more useful than any theoretical risk assessment of what the agent might do.