VaikoraVaikora

Vaikora › Blog › Product Updates

vaikora-guard-mcp: install it, write one rule, block your first tool call before it runs

Product Updates · Nov 22, 2026 · 4 min read

The gap in most MCP deployments is not authentication or authorization at the model layer. It is the absence of any deterministic check on what a tool call is about to do, at the moment it is about to do it. vaikora-guard-mcp closes that gap.

What the proxy actually does

vaikora-guard-mcp sits inline between an MCP client and the tool server it is calling. When an agent requests a tool call, the proxy intercepts the full request, evaluates it against your policy, and issues one of four decisions before the call reaches the tool: allow, block, modify, or require-approval. If the decision is block, the tool never runs. The agent receives a structured refusal. The whole exchange is written into a SHA-256 hash-chained audit log before the response travels back.

The binary is MIT-licensed, written in Python, and published on PyPI as `vaikora-guard-mcp`. It is self-hosted. You run it; your data does not leave your infrastructure.

Installing and wiring it up

```
pip install vaikora-guard-mcp
```

Point your MCP client at the proxy host instead of the tool server directly. The proxy forwards allowed calls transparently and returns structured policy decisions for everything else. No SDK migration. No changes to the tool server.

The first time you start it, you need a policy file. That can be as simple as a YAML block that names the tools you want to restrict and the conditions that trigger each decision.

A concrete scenario

Say your agent has access to a `write_file` tool and a `run_query` tool. You want to allow read-only SQL but block any query that contains a `DROP` or `DELETE` statement, and you want any `write_file` call targeting a path outside `/tmp` to require a human approval step before execution.

In your policy file, you define two rules. The first matches `run_query` calls where the query string contains `DROP` or `DELETE` and sets the decision to `block`. The second matches `write_file` calls where the `path` parameter does not start with `/tmp` and sets the decision to `require-approval`.

When the agent sends a `run_query` with `DELETE FROM users WHERE id=42`, the proxy evaluates the rule, issues a block decision at P50 8ms, writes the decision to the audit chain, and returns a structured error to the agent. The database never sees the query. When the agent sends `write_file` targeting `/etc/cron.d/agent-task`, the proxy holds the call and routes it to your approval queue. A human approves or denies it. Only then does the tool run.

That is deterministic policy enforcement. The same input always produces the same decision. There is no probabilistic judgment involved.

What is in the audit log

Every decision, allowed or blocked, is written as a signed record in a SHA-256 hash-chained log. Each record includes the tool name, the full parameters, the policy rule that matched, the decision issued, and a timestamp. Because the chain is hash-linked, you can verify after the fact that no record was modified or deleted. That property matters for compliance investigations. If you need to demonstrate to an auditor that your agent never executed a restricted action, the log is the evidence.

Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 ship with the enterprise control plane if you need rule sets mapped to those frameworks rather than writing them from scratch.

Where this fits in a larger enforcement stack

vaikora-guard-mcp is the open-source gateway component. The Vaikora Control Plane adds the management layer: centralized policy authoring, approval workflows, alerting, and integration with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub. If you are running multiple agents across different tools or teams, the control plane is where you manage policy consistently rather than maintaining separate YAML files per deployment.

For teams that want to start with the open-source proxy and add the control plane later, the architecture is the same. The gateway does not need to be replaced.

What to check next

Clone the repo or install from PyPI and run the proxy against a test tool server. Write one rule that blocks a call you would not want your agent to make in production. Verify that the audit log captures the decision correctly. That exercise will surface whether your current MCP setup has any visibility into tool call content at all, and whether you have anything in place to stop a call before it executes rather than detecting the effect afterward.

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

More from the Vaikora blog