Portkey (US) is an AI gateway that sits between the application and model providers. It centralizes routing, fallback, usage limits, guardrails and observability in a single panel, with an open-source edition for self-hosting and managed plans.
What it does
- Routes calls across OpenAI, Anthropic, Google, Azure, Bedrock and other providers
- Automatic fallback and retries when a provider fails or hits a limit
- Load balancing across keys and deployments of the same model
- Input and output guardrails, plus sensitive-data detection
- Logs, metrics and per-request cost in one panel
How it works
- The application calls an endpoint compatible with the OpenAI format
- The gateway applies policies, picks the provider and records the request
- Virtual keys grant access per project, team or user
- The open-source edition runs on your own infrastructure with Docker
Models and providers
- Catalog with hundreds of models from dozens of providers
- Support for text, image, audio and embedding models
- Providers can be added with your own key (BYOK)
Public pricing
- Open-source edition free for self-hosting
- Free Developer plan with a monthly log limit
- Paid Production plan by log volume
- Enterprise with SSO, VPC and data residency options
Strengths
- One panel for routing, cost and governance
- Open-source edition with no dependency on the managed service
- Guardrails and policies without changing application code
Points of attention
- Governance features vary between open source and paid plans
- Requires key and policy setup before production
- The extra layer adds a dependency point