Vaikora › Blog › Product Updates
What vaikora-guard-mcp actually does to an agent tool call: code, policy, and the MIT license explained
Publishing a proxy as open source under the MIT license means anyone can read the enforcement logic, audit the decision path, and run it without paying for anything. That is the point. Here is what the code actually does and why the design choices matter for teams building or securing AI agents.
What problem the proxy solves
An AI agent calls tools. Those tool calls might write to a database, read a file, POST to an external API, or delete a record. Without a gate in front of those calls, the only control is the agent's own reasoning, which is not deterministic and is not auditable.
vaikora-guard-mcp sits inline between the agent and the tools it calls via the Model Context Protocol. Before the tool call executes, the proxy evaluates it against policy and returns one of four decisions: allow, block, modify, or require-approval. The action either executes or it does not. There is no post-execution detection step.
How it is packaged
The package is published on PyPI as `vaikora-guard-mcp`. It is written in Python, MIT licensed, and self-hosted. You pull it, run it in your own environment, and it handles MCP traffic. No data leaves your infrastructure as part of the proxy operation. The MIT license means you can inspect, fork, or embed it without restriction.
The proxy ships six content inspection modules. Each module targets a specific category of risk: credential patterns, PII, instruction injection attempts, schema violations, and similar. When a tool call arrives, the modules run against the payload before any downstream execution happens.
What a policy check looks like in practice
Say an agent is operating under a HIPAA compliance preset. It receives a user message and decides to call a tool that writes records to a clinical database.
The tool call payload arrives at the proxy. The PII inspection module scans the arguments. It finds a pattern matching a patient identifier. The proxy checks that against the active policy for that agent session. The policy says: writes containing patient identifiers to this tool class require approval.
The proxy returns a `require-approval` decision. The tool call does not execute. The decision, the matched pattern, the policy rule that triggered, the timestamp, and the session context are all written to the audit log. That log entry is signed into a SHA-256 hash chain, so the record cannot be altered after the fact without breaking the chain.
If a human approver clears the action, the proxy forwards the original call. If not, the agent receives the block and the session continues without the write having happened.
The same mechanism works for `block` and `modify`. Modify is useful when a tool call is structurally valid but contains content that should be redacted before the downstream tool receives it. The proxy rewrites the argument, forwards the sanitized call, and logs both the original and the modified payload.
Determinism is the design constraint
Every enforcement decision in the proxy is deterministic. Identical input, same active policy, same decision every time. This matters for two reasons.
First, auditability. If a compliance auditor asks why a specific tool call was blocked at 14:32:07 on a given date, the answer is in the log and the decision can be reproduced from the inputs. There is no probabilistic model involved in the enforcement path.
Second, testing. Because the decision logic is deterministic and the code is open source, you can write unit tests against the policy engine before you deploy agents into production. You can verify that a payload matching a given pattern triggers the expected decision under a given policy configuration.
Connecting the proxy to the enterprise control plane
The open-source proxy is the enforcement layer. Policy authoring, session management, approval workflows, compliance preset configuration, and the audit log interface live in the Vaikora Control Plane, which is the commercial product hosted by Data443.
The two connect via a configuration the proxy fetches at startup and can refresh in a running session. If you want to run the proxy entirely offline against a static policy file, that is also supported. The MIT license does not restrict how you configure it.
For teams already using Microsoft Sentinel, SentinelOne, or CrowdStrike Falcon, the control plane pushes enforcement signals into those platforms natively. High-severity blocks appear as incidents or threat intelligence indicators in those tools without custom engineering.
What to check next
If you want to evaluate the enforcement logic before committing to any control plane, start with the PyPI package. Read the content inspection module code. Write a test payload that matches a pattern your environment cares about and verify the decision. The enforcement path is in the open.
For compliance preset details covering HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001, those are documented in the control plane configuration reference.
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