agnel.io

in build security and AI

Vetrix MCP Gateway

An MCP gateway that lets AI agents use only the tools their owners approved, and blocks a tool the moment its definition changes.

fingerprint pinned per tool
SHA-256
Kubernetes resources: MCPServer, MCPAccess
2
Go tests
34
Stack
  • Go
  • Kubernetes
  • MCP
  • OIDC

MCP servers describe their tools in plain text, and the model trusts that text. A server can ship a harmless description, get approved, and later change it into an instruction to leak secrets. Most teams have no control over which servers and tools their agents can reach.

The gateway pins every tool to a fingerprint of its full definition. Owners approve tools in Git, where the reviewer reads exactly what the model will read, and new servers start disabled. At run time the gateway checks each caller’s OIDC token, applies per-team access grants that deny by default, hides tools nobody approved, and refuses any call to a tool whose definition no longer matches.

Client credentials are never forwarded upstream, and every decision lands in an audit log.

vtx · type help, Tab to complete, Esc to close