A combined LLM and MCP gateway that scopes tool access and signs every decision

Verosek was founded in 2025 by Vats Shah and is based in Mumbai, operating without outside funding. Its premise is that instructing an agent to behave is not a control, because the enforcement point that matters is the tool boundary, where a call actually reaches a system of record. The gateway is adopted by changing a base URL: it presents the same interface as the OpenAI, Anthropic, and Gemini SDKs while routing to more than fifteen model endpoints behind it, which puts every model call and every tool execution on a single inspectable path without requiring application code to be rewritten around a new client.

Four layers sit on that path. The AI Gateway handles routing and failover. Shield runs a set of twenty-four checks over requests, responses, and post-tool output for prompt injection, personally identifiable information, and off-topic drift, processing offline rather than calling out to a judging model, and returning a deterministic allow, block, modify, or hold verdict. MCP Tools governs what an agent may invoke across a library of connectors, with permissions scoped per key and per argument so a call can be narrowed, masked, capped, or refused before it lands. Audit Vault records each decision as an HMAC-SHA256 signed receipt that can be verified offline.