The gap in most AI agent deployments is timing. Policy checks happen after a tool call runs, in logs, in SIEMs, in post-incident reviews. By then the action is done: the file was written, the API was called, the data was exfiltrated. The question worth asking before you deploy an agent into production is whether you can move that check to before execution, without rewriting your existing agent code.
That is what the vaikora-guard-mcp project does. It is a small Python proxy, MIT licensed, published on PyPI, that sits between an MCP client and the LLM or tools it calls. Every tool call goes through a policy evaluation before it runs. The result is one of four outcomes: allow, block, modify, or require-approval. Nothing executes until the policy engine returns a decision.
What "inline" Actually Means Here
In a standard MCP setup, the client sends a tool call and the server executes it. Vaikora Guard MCP inserts itself into that path as a proxy. The client talks to the proxy. The proxy evaluates the call against configured policy. If the call is allowed, the proxy forwards it to the actual MCP server. If it is blocked, the call never reaches the server.
The enforcement runs at P50 8ms, P99 45ms, so the latency added to any individual call is small. At scale the proxy handles 10,000 or more actions per second.
Critically, the enforcement is deterministic. Identical input always produces identical output. There is no probabilistic scoring that might return different results on the same call in different contexts. If a policy says "block tool calls that include PII in the arguments," that rule fires every time, not most of the time.
How Policy Is Configured
Policy is defined in a configuration file, not in agent code. That separation matters: the agent developer and the security team do not need to coordinate on every rule change. When a compliance requirement changes, the policy file changes. The agent itself does not.
A concrete example: an agent that queries a customer database and writes results to a file system. Without a policy gate, nothing stops that agent from writing to an arbitrary path, or from including raw PII in a log file. With Vaikora Guard MCP in place, you define rules that restrict write operations to approved directories and that inspect argument content for PII patterns before allowing the call. Those rules enforce at every invocation, not just when someone remembers to review a log.
Vaikora Guard MCP ships with compliance presets for HIPAA, PCI-DSS, GDPR, SOC 2, and ISO 27001. For regulated environments, those presets give you a starting configuration that reflects the data handling requirements of each standard, rather than building rules from scratch.
The Audit Chain
Every decision the proxy makes is written to a SHA-256 hash-chained, tamper-evident audit log. Each entry references the previous one, so any modification to a past record breaks the chain and is detectable.
For security operations and compliance purposes, this matters more than the enforcement itself in some cases. Enforcement prevents the bad action. The audit chain proves, to an auditor or an investigator, what the system evaluated, what decision it made, and in what order. That ordered record is not reconstructable from application logs after the fact.
Installing and Running It
The package is on PyPI as `vaikora-guard-mcp`. Installation is a standard `pip install vaikora-guard-mcp`. The proxy is self-hosted. Your data does not leave your environment during policy evaluation.
The commercial control plane, managed by Data443, adds centralized policy management, dashboards, and SIEM integrations including Microsoft Sentinel, SentinelOne, CrowdStrike Falcon, and AWS Security Hub. But the enforcement proxy itself runs in your infrastructure under the MIT license.
For teams already using MCP-compatible agents, there is no SDK migration required. The proxy is compatible with 12 LLM providers as configured.
What to Check Before You Deploy
Before putting any AI agent into production, verify three things about your current setup.
- Where does policy evaluation happen relative to execution? If the answer is "after," you are relying on detection and rollback rather than prevention.
- Are your tool call policies deterministic? If a rule applies inconsistently, it is not a guardrail, it is a suggestion.
- Do you have an ordered, tamper-evident record of every policy decision? Logs that can be modified after the fact do not satisfy most audit requirements.
If any of those checks fails, the place to start is the vaikora-guard-mcp repository on PyPI. Read the configuration format, run it against a non-production agent, and verify that the audit log structure matches what your compliance program requires before you scale the deployment.