The heaviest LLMOps surface in one place: routing, caching, 40+ guardrails, and observability behind a single gateway, fully open source since March 2026.
What it is
Portkey is an AI gateway spanning 1600+ models across 250+ providers, with conditional routing, circuit breakers, semantic caching, 40+ pre-built guardrails, and an MCP gateway for tool access. Where a thin proxy just translates and forwards, Portkey folds the whole cross-cutting stack, cost, safety, and tracing, into one control plane. It went fully open source under Apache 2.0 in March 2026, so you can self-host the same surface the cloud runs.
The one job
It is the one gateway that does all four at once, so you stop stitching a router, a cache, a guardrail layer, and an observability tool into each other and instead configure them as one thing at the edge.
Reach for it when
Skip it when
A minimal look
import OpenAI from "openai";
// point the OpenAI client at Portkey; policy travels in the headers
const client = new OpenAI({
baseURL: "https://api.portkey.ai/v1",
apiKey: process.env.OPENAI_API_KEY,
defaultHeaders: {
"x-portkey-api-key": process.env.PORTKEY_KEY,
"x-portkey-config": process.env.PORTKEY_CONFIG, // routing + cache + guardrails
},
});
The principle it teaches
Consolidate cross-cutting concerns, cost, safety, and tracing, at the gateway rather than in app code. Portkey is a single tool that touches Weeks 3, 4, and 5 at once, which is exactly why the boundary is the right place for them.
Where it fits your labs
This spans Weeks 3 to 5. The cost work of Week 3, the safety and guardrail work of Week 4, and the observability of Week 5 all have a home in this one gateway, so you can see how a real control plane earns its keep.