Most MCP servers have no policy layer. A tool call comes in, the server executes it, and whatever happened, happened. If you want to know whether the action was appropriate, you find out afterward, in a log, or not at all.
Vaikora Guard MCP is a small, MIT-licensed Python proxy published on PyPI as `vaikora-guard-mcp`. It sits between an AI agent and the tools it calls, checks every action against a policy definition before execution, and returns an allow, block, modify, or require-approval decision. The tool call either proceeds, is changed, or is stopped. No action runs first and gets reviewed second.
What the Proxy Actually Does
The proxy intercepts tool call requests at the MCP layer. Before forwarding a call to the target tool, it evaluates the call against your policy configuration. The evaluation is deterministic: identical input produces an identical decision every time. That matters for auditability, because a policy that produces inconsistent outputs cannot be tested or reasoned about reliably.
Every decision is written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so you get a tamper-evident, ordered record of what the agent tried to do, what the policy said, and what actually ran.
Enforcement latency runs at a P50 of 8ms and a P99 of 45ms, so the proxy does not materially slow down agent workflows.
A Concrete Example
Say you have an agent with access to a filesystem tool and a database query tool. Without a policy layer, the agent can read or write any path the process user has access to, and can run any query the database credentials allow. Your IAM and RBAC controls set the outer boundary, but nothing inside that boundary checks whether a specific action is appropriate for the task at hand.
With Vaikora Guard MCP in front of those tools, you write a policy that permits filesystem reads under `/data/reports/` and blocks writes entirely during the current session. You add a rule that requires human approval before any query that includes a `DROP` or `DELETE` keyword. The proxy enforces both before execution. The agent never receives a success response from a blocked call, and the approval gate pauses execution until a human responds.
The agent sees a policy rejection or an approval prompt. The IAM and RBAC roles stay in place underneath. Vaikora Guard MCP adds a runtime gate on top, not a replacement for existing access controls.
Installing and Configuring It
The package is on PyPI:
“`
pip install vaikora-guard-mcp
“`
Configuration is a policy file that maps tool names to rules. Rules specify conditions (parameter patterns, value ranges, keyword matches) and outcomes (allow, block, modify, require-approval). The policy file is evaluated at request time, so changes take effect without restarting the proxy.
The MCP-compatible design means the proxy works with any agent or orchestration framework that speaks MCP. No SDK migration, no changes to the agent code beyond pointing it at the proxy endpoint.
What You Get at the Governance Level
The audit log is the main artifact for compliance and investigation workflows. Because each entry is hash-chained, you can demonstrate to an auditor that the log has not been altered after the fact. That matters under frameworks like SOC 2, HIPAA, and GDPR, where you need to show not only what happened but that the record of what happened is reliable.
The open-source, self-hosted design means the log stays in your environment. You are not sending agent telemetry to a third party to get a policy decision.
For teams that need a control plane on top of the proxy, managing policies across multiple agents, tracking decisions in aggregate, and routing approval requests, the Vaikora Control Plane is Data443’s commercial offering. The open-source proxy connects to it, but it is not required to use the proxy.
What to Check Next
Pull the package from PyPI and read the policy configuration reference in the repository before writing your first rule. Start with the tool that carries the highest blast radius if misused, usually anything with write, delete, or external network access, and define a minimal allow policy before adding complexity. Run a few agent sessions, review the audit log, and confirm the decisions match your intent. That gap between what you expected and what the log shows is where policy tuning needs to happen.