Portkey

Portkey is an AI gateway: a control plane that sits between an organization's applications and agents and the many AI models they call. Rather than each application integrating directly with a dozen model providers, traffic flows through Portkey, which gives teams a single place to route requests, enforce policy, and see what their AI is actually doing. In 2026 Portkey was acquired by Palo Alto Networks, which positioned it as the AI Gateway for its Prisma AIRS security platform, reflecting how central this control point has become as enterprises move from chatbots to autonomous agents.

The gateway exposes one unified interface to a very large catalog of models across hundreds of providers, from OpenAI, Anthropic, Google, and AWS Bedrock to open-source and self-hosted models, so an application can switch or mix models without rewriting its integration. Around that routing layer Portkey adds the reliability features production systems need: automatic retries, fallbacks between providers, load balancing, and circuit breakers, plus semantic caching to cut repeated cost and latency.

Its second pillar is observability and governance. Portkey logs every request and response with token usage, cost, and latency, giving teams visibility into model and tool interactions in one place, and notably treats Model Context Protocol tool calls and model calls as a single observable surface as agents take on more work. On top of that visibility it enforces control: rate limits, budget and cost guardrails to prevent runaway spend, access controls and virtual keys, and policy checks that can flag or block non-compliant behavior in real time. It also includes prompt management for versioning and reusing prompts.

Because autonomous agents act like highly privileged insiders, making many automated decisions across internal and external systems, this combination of routing, reliability, and runtime governance is exactly the gap Portkey targets, and the reason it became the foundation of a major security platform. The core gateway is available as open source and can be self-hosted, while enterprises increasingly consume it as part of a broader AI security and control stack for operating agents safely at scale.