Vaikora › Blog › Product Updates

vaikora-guard-mcp is MIT licensed and on PyPI: here is what it actually does at the tool-call layer

Product Updates · Nov 7, 2026 · 4 min read

When you give an AI agent access to tools through MCP, the protocol itself does not decide what the agent is allowed to call. That decision falls to whatever is sitting between the agent and the server. If nothing is sitting there, the agent runs whatever it requests.

vaikora-guard-mcp is an open-source proxy, MIT licensed, written in Python, and published on PyPI as `vaikora-guard-mcp`. It installs in a single command and places a policy enforcement point directly in the MCP request path, before any tool call executes.

What the proxy actually does

When an AI agent issues a tool call, the request hits vaikora-guard-mcp before it reaches the MCP server. The proxy evaluates the call against a deterministic policy and returns one of four decisions: allow, block, modify, or require-approval. The underlying tool never executes until the policy produces a result.

Identical input always produces an identical decision. There is no probabilistic scoring or model-based judgment at the enforcement layer. If a policy says block calls to `delete_record` from agents without an approval token, that call is blocked every time, without exception.

Every decision is written into a SHA-256 hash-chained, tamper-evident audit log. The chain gives you an ordered, verifiable record of what was allowed and what was stopped, which matters when a compliance auditor or incident investigator asks exactly what happened at 14:32 on a given date.

Installation and what you get out of the box

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

The package ships six content inspection modules that evaluate tool call payloads for sensitive data patterns, policy violations, and scope mismatches. It includes adapters for the major hosted model APIs and native MCP support.

Enforcement latency runs at P50 8ms and P99 45ms, so the proxy is not a meaningful bottleneck in normal agent workflows. At 10,000-plus actions per second, it holds up under parallel agent load without queuing.

Compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001 ship with the package, so you do not have to write policy from scratch for regulated environments. You load a preset, review what it blocks by default, and adjust from there.

A concrete scenario

Say you have an agent that queries a customer database via an MCP tool called `run_sql`. Without a gate in the path, the agent can call `run_sql` with any query it generates. With vaikora-guard-mcp sitting inline:

• A `SELECT` on non-PII columns passes the HIPAA preset and executes normally.
• A `SELECT` that pulls `ssn`, `dob`, or `diagnosis` fields triggers the PII pattern inspection module and is blocked or held for approval, depending on your policy configuration.
• Any `DROP` or `DELETE` statement hits the destructive-action rule and is blocked outright.
• The agent receives the block decision, the reason, and can be configured to request human approval before retrying.

The SQL never reached the database for the blocked cases. The audit log shows each decision in sequence, hashed and ordered.

Where the open-source boundary is

The proxy itself, vaikora-guard-mcp, is the open-source component. Self-hosted, no license key, no call-home requirement. You run it in your own environment and own the audit log it produces.

The enterprise control plane is the commercial product from Data443. It adds centralized policy management across multiple agents and environments, SIEM integrations with Microsoft Sentinel, CrowdStrike Falcon, SentinelOne, and AWS Security Hub, plus severity-mapped threat indicator pushing and Azure Security Center incident promotion. Those integrations are not in the open-source package.

If you are running a single agent in a controlled environment and want enforcement without infrastructure overhead, the open-source proxy covers it. If you are coordinating policy across a fleet of agents and need compliance reporting with a full audit trail pushed into your existing security tooling, that is what the control plane adds.

What to check next

Pull the package from PyPI, run it against a test MCP server, and read the audit log it produces after a blocked call. The log format is the clearest way to understand what the proxy actually records. If you are working in a regulated environment, load the relevant compliance preset first and review which tool call patterns it blocks by default before connecting it to production agents.

The source is on GitHub if you want to inspect the content inspection modules or contribute a policy rule.

More from the Vaikora blog