Architecture
The gateway runs as a sidecar in front of LLM traffic, an inline Python or Node.js SDK embedded in the application, or a hosted endpoint via the Vaikora Control Plane. Every proposed action is evaluated against the policy engine; the engine returns a decision (ALLOW, ALLOW_LOG, CONSTRAIN, BLOCK) plus a receipt id. Decisions land in under 500ms at the p95.
Adapters
First-class adapters for OpenAI, Anthropic, Google Gemini, and OpenRouter at the LLM level. Agent-to-agent (A2A) and Model Context Protocol (MCP) at the protocol level. The vaikora-guard-mcp server is registered in the Anthropic MCP registry and installable from PyPI.
Content modules
Six modules ship by default: PII detection, jailbreak detection, prompt injection detection, semantic risk classification, domain risk scoring, and email classification. Each module is independently configurable and replaceable; the engine composes their verdicts into a single policy decision.
License
MIT license on the gateway and the MCP server. Source on GitHub, contributions welcome. Production Control Plane is commercial.