15 posts tagged with "AI"

The biggest STOA release yet — 1,091 commits, 297 features, and a fundamental shift toward AI-native API management.

AI gateways require specialized rate limiting approaches that account for token consumption, streaming responses, and variable request costs. Traditional request-per-second limits fail to capture the true resource usage of AI workloads. This guide covers token-aware rate limiting strategies, per-tenant quota management, and implementation patterns for production AI gateways.

STOA Platform secures AI agent API access through five independent layers: mTLS certificate binding, OAuth 2.1 with PKCE, OPA policy evaluation, AI guardrails, and immutable audit logging. Each layer addresses a distinct threat class. Compromise of any single layer does not grant unauthorized access. This article describes the security architecture, threat model, and design rationale for each layer.

Custom MCP tools let you expose any API as an AI-native interface that Claude and other AI agents can discover and invoke automatically. This tutorial walks you through creating, registering, and testing a custom MCP tool using the STOA gateway, from initial YAML definition to live invocation by an AI agent.

Proxy throughput benchmarks tell you how fast a gateway can forward HTTP requests. They tell you nothing about whether that gateway can serve AI agents, enforce guardrails on tool calls, or govern autonomous sessions. We built a new benchmark that measures what actually matters.

AI agents need programmatic API access, but traditional authentication patterns designed for human users — browser cookies, session tokens, OAuth2 authorization code flows — don't work. AI agents are autonomous services, not users. They operate without browsers, without human-in-the-loop interactions, and at machine speed. This article presents five authentication patterns that work for AI agents, from the simplest (API keys) to the most secure (mTLS certificate binding), with practical implementation examples for each.

This is part of the What is an MCP Gateway series. For the broader context on why AI agents need specialized infrastructure, see Connecting AI Agents to Enterprise APIs.

A single developer shipping 72 story points per day across 7 components, 22 PRs per week, with zero regressions on main. This is not a theoretical exercise — it is the measured output of STOA Platform's AI Factory during Cycle 7 (February 9-15, 2026). This article explains the architecture, the coordination protocols, and the hard lessons that make it work.

If you are building an MCP gateway or any complex open-source platform, the patterns described here are directly reusable. They are not tied to STOA — we extracted them into a reusable pattern library (HEGEMON) that any project can adopt.

Kong and STOA both support the Model Context Protocol, but they approach it from opposite directions. Kong added MCP via plugins on its proven Nginx/Lua stack. STOA built MCP into the gateway core from day one. This article compares the two specifically on MCP capabilities — tool discovery, transport, authentication, governance, and agent workflow support — so you can choose the right MCP gateway for your AI agent architecture.

Every REST API endpoint can become an MCP tool that AI agents discover and invoke automatically. The conversion is a mapping exercise: your OpenAPI spec already contains the tool name, description, parameters, and endpoint URL that MCP needs. This guide walks through the process from a single endpoint to bulk automation using ToolSet CRDs.

The Model Context Protocol (MCP) is a JSON-RPC 2.0 based protocol that standardizes how AI agents discover, authenticate with, and invoke external tools. It defines four phases — initialization, discovery, invocation, and streaming — over pluggable transports including SSE, WebSocket, and stdio. This article covers the protocol internals that matter for production deployments.