Vaikora › Blog › Product Updates
What vaikora-guard-mcp actually does when an agent tries to call a tool: a code-level walkthrough
Every tool call an AI agent makes is a decision point. Without a gate at that point, the agent either executes or it does not, based entirely on whatever the LLM decided. vaikora-guard-mcp puts a policy check at that exact moment, before execution, using code you can read, run, and modify yourself.
The package is MIT-licensed, written in Python, and published on PyPI as `vaikora-guard-mcp`. That means you can inspect every line before trusting it with agent traffic.
What the proxy intercepts
vaikora-guard-mcp sits between an AI agent and the MCP server it talks to. When the agent issues a tool call, the request goes through the proxy first. The proxy evaluates it against configured policy and returns one of four decisions: allow, block, modify, or require-approval.
Only if the decision is `allow` does the tool call proceed. Block stops it immediately and returns a structured error. Modify rewrites the request before forwarding it. Require-approval holds the call and waits for a human signal before continuing.
This is not a log-and-alert pattern. The check happens inline, before the tool function runs.
What "deterministic" means in practice
Deterministic enforcement means identical input always produces identical output. There is no model scoring, no confidence interval, no threshold to tune. If a tool call matches a policy rule, the outcome is the same every time, regardless of time of day, load, or prior context.
A concrete example: suppose your policy says "block any file-write call where the destination path matches `/etc/*`". An agent that generates a call to `write_file("/etc/cron.d/malicious", ...)` will be blocked. Every time. The policy is a function, not a prediction.
This matters for auditability. Every decision vaikora-guard-mcp records goes into a SHA-256 hash-chained audit log. Each entry is signed into the chain, so the record of what was allowed or blocked cannot be altered without breaking the chain. That is the kind of evidence a compliance review or incident investigation can actually use.
Installing and wiring it in
```
pip install vaikora-guard-mcp
```
Configuration is a YAML policy file. You define rules against tool name, argument patterns, calling identity, and content inspection results. The proxy loads the file at startup. Rule changes take effect on reload without redeployment.
The proxy handles MCP-compatible traffic, so any agent already using MCP does not need an SDK migration. You point the agent at the proxy endpoint instead of the MCP server, and the proxy forwards approved calls onward.
Enforcement latency at P50 is 8ms, P99 is 45ms. For most agent workflows that is below the threshold where users or downstream systems notice the difference.
What the six content inspection modules do
The gateway ships six content inspection modules that run as part of the evaluation chain:
- PII detection: flags calls containing personal identifiers in arguments
- Secrets detection: catches API keys, tokens, and credentials in outbound payloads
- Prompt injection detection: identifies instruction-override attempts embedded in tool arguments
- Data classification: evaluates content against sensitivity labels before allowing output
- URL and domain inspection: checks destinations against threat intelligence feeds
- Custom regex and pattern rules: catch organization-specific data patterns
Each module can be toggled and configured independently. You can run all six or only the ones relevant to your environment.
Connecting the open-source proxy to the enterprise control plane
vaikora-guard-mcp is the self-hosted enforcement layer. The Vaikora Control Plane, hosted by Data443, is the commercial product that adds centralized policy management, cross-agent audit aggregation, SIEM integration, and compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001.
The two are designed to work together, but the proxy runs independently if you want to start with local policy files and no cloud dependency. That is a reasonable starting point for a proof of concept or a security review before committing to the full platform.
What to check next
If you are evaluating this for a real deployment, the place to start is the policy file schema. Review what rule conditions are available, map them against the tool calls your agents actually make today, and check whether the four decision types cover the enforcement gaps you have. The source is on PyPI and the license is MIT, so there is no barrier to reading the inspection module code before you run it in any environment that matters.
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