Portkey

Portkey is an AI gateway and operations layer for applications that call multiple language-model providers, agents, and MCP tools. Instead of wiring every application directly to each vendor, a team can route requests through Portkey, apply policy, collect traces and cost data, configure fallbacks, manage prompts, and centralize credentials. The fetched vendor metadata and search index explicitly list AI Gateway, Observability, Guardrails, Prompt Engineering Studio, Agents, MCP Gateway, role-based access control, network-level guardrails, and organization-wide audit logs. That breadth differentiates it from a simple model marketplace. Compare the gateway layer with Cloudflare AI Gateway, and compare its telemetry with Langfuse or Helicone.

Existing staged vendor data records Production at $49 per month, with overages of $9 per additional 100,000 requests, and Enterprise at custom pricing. During the September 8, 2026 research, the vendor pages returned large script-heavy Framer HTML; product labels were recoverable, but visible pricing content could not be independently extracted from the fetched page. Those figures are therefore retained with a manual-verification flag. Buyers should confirm whether the base price includes team seats, retention, guardrails, prompt management, MCP controls, support, and a specific request allowance. Model-provider charges are separate from gateway fees, so total cost equals Portkey fees plus the underlying model and any vector, tool, or hosting services.

Portkey is most useful when an application needs reliability or governance across providers. A production chat service can retry transient failures, route to a fallback model, enforce allowed providers by team, attach metadata, and compare latency, token usage, and cost. A platform group can issue virtual keys without distributing provider credentials. An agent team can put authentication, policy, and observability in front of MCP servers. The practical evaluation should replay representative traffic through direct and gateway paths. Measure p50 and p95 added latency, successful fallback rate, trace completeness, cost attribution, cache correctness, and behavior during provider outages. The production agent monitoring guide provides a useful checklist.

The major benefits are one integration surface, portable routing logic, centralized logs, and guardrails that do not have to be rebuilt in every application. The costs are another dependency in the request path and the need to understand how sensitive prompts and outputs are stored. A gateway outage can affect every connected model. Misconfigured retries can multiply spend or repeat side effects. Caching can return stale or inappropriately shared content. Guardrails can block legitimate requests or miss unsafe ones. Require scoped keys, data-retention controls, redaction tests, regional requirements, idempotency, rate limits, exportable traces, and a documented bypass or failover plan.

The honest fit is a team with several providers, several applications, or real compliance and reliability requirements. A prototype using one model may not need Portkey yet. Before committing, calculate monthly request volume under the published overage unit, run an outage exercise, and compare managed convenience against operating an open-source proxy. Portkey can simplify a fragmented AI stack, but only if the additional hop produces measurable control and reliability.