Vaikora › Blog › Product Updates
MIT-licensed, self-hosted, MCP-compatible: what Vaikora Guard MCP enforces and how to run it in under ten minutes
Shipping an AI agent without a policy gate on its tool calls is equivalent to giving a junior employee root access and trusting they will figure out the limits themselves. Vaikora Guard MCP is the open-source answer: a Python proxy, published on PyPI as `vaikora-guard-mcp`, that sits inline between an MCP client and every tool it invokes, and enforces an allow, block, modify, or require-approval decision before execution.
What the package is and what it is not
`vaikora-guard-mcp` is MIT-licensed, written in Python, and self-hostable. You install it in your own environment. No data leaves to a Data443 endpoint unless you separately connect the enterprise Control Plane. The proxy is MCP-compatible, meaning it speaks the same wire protocol as any existing MCP server, so there is no SDK migration and no rewrite of the agent.
It is not a monitoring wrapper. The enforcement decision happens before the tool call completes. If a call is blocked, the tool never executes. If a call requires approval, execution pauses until a human responds. Detection after the fact is not the model here.
How the enforcement path works
The proxy intercepts each tool call, extracts the method name, arguments, and any metadata, then evaluates that payload against a deterministic policy set. Identical input always produces an identical decision. That is not aspirational: it is structural. There is no probabilistic scoring that changes run to run.
The six content inspection modules shipped with the gateway cover categories like data exfiltration patterns, credential exposure, and calls to known-malicious infrastructure. Policy authors write rules against these categories. The result for each call is one of four outcomes:
- Allow: the call proceeds immediately.
- Block: the proxy returns an error to the agent; the tool never sees the request.
- Modify: the proxy rewrites the call, for example stripping a field that would leak PII, then forwards the sanitized version.
- Require-approval: the call is queued; a human reviewer sees the pending action and approves or rejects it in real time.
Every decision is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry is signed into the chain, so any post-hoc modification of a record breaks the chain and is detectable.
A concrete installation scenario
Say you are running an MCP-compatible coding agent that has access to a file-read tool and an HTTP fetch tool. Without a policy layer, the agent can read any file the process user can access and fetch any URL. That is the attack surface.
With `vaikora-guard-mcp` installed as the proxy:
- `pip install vaikora-guard-mcp`
- Point the proxy at your existing MCP server endpoint.
- Load a policy file that blocks reads outside `/workspace/`, blocks HTTP fetches to non-allowlisted domains, and requires approval for any call that matches a credential pattern in its arguments.
Now the agent operates within a policy boundary. A prompt injection that attempts to exfiltrate `/etc/passwd` hits the block rule. A tool call to an attacker-controlled URL hits the domain allowlist rule. Neither executes. The audit log records both attempts with full argument payloads.
Latency and throughput at the enforcement layer
A common concern with inline enforcement is added latency. At P50 the enforcement decision adds 8ms. At P99 it adds 45ms. For agentic workloads where individual tool calls often involve network I/O measured in hundreds of milliseconds, that overhead is not the bottleneck. The gateway supports 10,000 or more enforcement decisions per second, so throughput is not a constraint for most deployment scales either.
Connecting to the enterprise Control Plane
The open-source proxy enforces policy locally. If your team needs centralized policy management across multiple agents, audit log aggregation, approval workflows with role-based reviewers, or compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, or ISO 27001, the Vaikora Control Plane is the commercial layer that sits above the proxy. The gateway is the enforcement point; the Control Plane is the management surface.
The proxy can also push signals into Microsoft Sentinel, SentinelOne, CrowdStrike Falcon, or AWS Security Hub if you want AI agent threats visible inside your existing security operations tooling. That integration is configured at the Control Plane level, not the open-source proxy level.
What to check next
Install the package from PyPI and review the default policy file before you configure anything. The policy file is the most important artifact: rules that are too permissive defeat the point, and rules that are too restrictive will block legitimate agent behavior. Start with a dry-run mode if one is available in your version, capture a sample of real tool calls your agent makes, then write rules against that baseline rather than hypotheticals.
The source is on GitHub under the MIT license. Read the enforcement decision logic directly. For a proxy that sits between your agent and production tools, understanding exactly what the code does on each call is worth the hour.
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