Portkey: AI Gateway with Guardrails

Overview

Portkey is a production AI gateway that handles routing, observability, and guardrails for LLM traffic. It supports 250+ models across providers and sits between application code and the model APIs, providing a control plane for managing production LLM workloads.

Palo Alto Networks announced its intent to acquire Portkey in April 2026 -- a signal that AI gateway capabilities are now considered security-critical infrastructure, not just developer tooling.

Pricing

Portkey charges per recorded log -- you pay your LLM providers separately for actual model usage. A free tier is available for initial development. Pro and Enterprise tiers unlock longer retention, compliance features (SOC 2, GDPR), and dedicated support. Enterprise adds SSO/SCIM, audit logging, and custom data residency.

The open-source gateway (MIT license) is available on GitHub for self-hosting with no licensing fee.

Key capabilities

Guardrails: Request and response filters including jailbreak detection, PII redaction, and policy-based enforcement. This is the feature that distinguishes Portkey from passthrough gateways like OpenRouter.

MCP Gateway: A centralized control plane for managing Model Context Protocol servers across an organization -- teams can route, audit, and policy-enforce MCP traffic alongside standard LLM requests.

Observability: Detailed logs with latency, cost, token counts by app, team, and model. Direct competitors are Helicone and Langfuse.

Workspaces: Role-based access, separate API key sets per team, and audit trails -- enterprise multi-tenancy built in.

Where it fits

Portkey is the right choice when guardrails and governance are requirements, not nice-to-haves. It handles the security and compliance layer that pure inference aggregators don't address. For teams that need PII redaction before prompts reach any provider, or policy enforcement on what models agents can call, Portkey covers that at the gateway layer.

The MCP Gateway capability is the newest differentiator (2026): as agent frameworks that use MCP proliferate, controlling what MCP servers can be invoked and auditing their usage becomes an operational necessity.

Field notes

• Palo Alto Networks announced acquisition intent for Portkey in April 2026, citing the rise of AI agents as the driver. The press release specifically called out MCP Gateway as the strategic asset: PANW wants to enforce AI agent security policies at the gateway layer, not at the application layer. [dev-post, 2026-04-15]

See also

OpenRouter, Vercel AI Gateway, Helicone, Langfuse