Policy enforcement for AI agents has a timing problem. Most approaches check what an agent did after the action completed. vaikora-guard-mcp sits inline and checks what the agent is about to do, before the tool call executes.
The package is MIT-licensed, written in Python, and published on PyPI as `vaikora-guard-mcp`. You can read the source, fork it, run it on your own infrastructure, and inspect exactly what policy is applied at each step.
What the proxy actually does
vaikora-guard-mcp acts as an MCP-compatible proxy between your AI agent and the tools it calls. Every tool invocation passes through the proxy. The proxy evaluates the call against a configured policy and returns one of four decisions: allow, block, modify, or require-approval. The tool only executes if the decision is "allow" or "modify." Nothing executes while the proxy is waiting for an approval decision.
That ordering matters. A block decision means the tool never ran. There is no rollback step, no compensating transaction, no incident to investigate after the fact.
Every decision, including allows, is written to a SHA-256 hash-chained audit log. Each entry is signed into the chain, so you have a tamper-evident, ordered record of every policy outcome your agent produced.
A concrete scenario
Say your agent is wired up to a file-write tool and a SQL execution tool. Without a policy gate, the agent calls both tools based on whatever the model decides to do next. With vaikora-guard-mcp in place, you can write a policy that blocks any SQL execution containing a `DROP` or `TRUNCATE` statement, requires human approval before any file write to a path under `/prod`, and allows everything else to pass.
When the agent issues a `DROP TABLE` call at 2 a.m., the proxy blocks it. The audit log records the blocked call with a timestamp and the policy rule that matched. The database is untouched. No one needs to restore a backup.
That same configuration applies deterministically. Identical input, identical decision, every time. There is no model judgment involved in the enforcement step.
How to install and configure it
“`
pip install vaikora-guard-mcp
“`
Configuration is a policy file you write and control. The proxy reads it at startup. You point your MCP-compatible agent at the proxy endpoint instead of directly at the tool server. No SDK migration is required.
The self-hosted proxy handles enforcement. If you want a control plane with a UI, policy management across multiple agents, compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001, and native integrations with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub, that is the Vaikora commercial product. The proxy and the control plane are separate. You can run the proxy standalone indefinitely.
What "open source" means for auditability
Publishing under MIT means your security team can read exactly what the proxy does to a tool call. For regulated environments, that matters. You are not trusting a black box to enforce policy on your AI agent. You can audit the enforcement logic itself.
The hash-chained audit log is also inspectable. Each record includes the tool name, the input parameters, the policy decision, the matched rule, and the timestamp. Because each entry is linked to the previous one by hash, you can detect any attempt to delete or alter a log entry.
That log format is what compliance reviewers and auditors will want to see when they ask how you are controlling AI agent behavior in production.
Who should run this
If you are running any AI agent that calls external tools, writes to storage, executes queries, or triggers APIs, and you do not have a gate on those calls, you have a control gap. The gap is not theoretical. Agents mis-step in production. Models hallucinate tool arguments. Prompt injection causes agents to call tools they should not.
vaikora-guard-mcp does not prevent bad model outputs. It prevents bad model outputs from completing as tool executions.
What to do next
Pull the package, read the source on PyPI, and wire it in front of one tool in a test environment. Write a policy that blocks something you would not want an agent to do in production. Verify that the blocked call appears in the audit log with the correct rule reference.
If you need policy management across more than a handful of agents, or you need the SIEM and EDR integrations, check the Vaikora Control Plane documentation on the Data443 site. The proxy is the starting point. The control plane is what scales it.