Portkey.ai

Portkey.ai builds an AI gateway that sits between applications and model providers, combining request routing, observability, reliability, governance and cost management. Palo Alto Networks completed its acquisition of Portkey on May 29, 2026, establishing it as the core AI Gateway for Prisma AIRS. The acquirer outlined planned integrations with runtime security, agent identity security through Idira and AI observability through Chronosphere.[5][7]

Before the acquisition, Portkey developed tools for operating LLM applications in production. Its founders’ earlier application work provides context for the initial platform and its subsequent expansion into guardrails and agent infrastructure.[3][4][6]

Rohit Agarwal and Ayush Garg founded Portkey.ai in January 2023. In their August 2023 founders’ account, they recalled encountering increasing latency, missing request and error visibility, abuse, rate limits, poor key management and client-data privacy complications while scaling earlier GPT-3 applications. Their distinction was practical: “building a prototype on top of LLMs is a very different problem from scaling them to production.”[2][3]

The initial full-stack LLMOps platform combined an observability layer, AI Gateway, prompt manager, experimentation and evaluations framework, and security and compliance protocols. Its integration example changed an OpenAI client’s baseURL and added Portkey headers, placing the gateway in the request path and enabling automatic logging. The hosted gateway balanced requests across API keys and model providers to address uptime, errors and rate limiting.[3]

Springworks CEO Kartik Mandaville described a concrete use case: its Albus employee help desk was handling more than 10,000 questions daily, and costs, latency and OpenAI rate limiting were becoming difficult to manage. In a testimonial carried by Portkey, he credited its analytics and semantic caching with helping address those problems.[2][3]

On August 23, 2023, Portkey announced it was raising a $3 million seed round, led by Dev Khare and Manjot Pahwa from Lightspeed India. It reported having worked with more than 100 companies, serving millions of requests daily and processing more than 50 million LLM requests cumulatively. Its announced research plans included real-time canary testing for continuous model and prompt evaluation, and fine-tuning models.[3][2]

On August 14, 2024, Agarwal and Garg announced guardrail integration in the open-source gateway and free availability on Portkey’s hosted application. They distinguished robust API operations and request/response transformations from the underlying model’s behavior: making requests more reliable did not resolve inaccurate, biased or harmful outputs. Guardrails added another decision point, allowing verdicts on inputs or outputs to direct request orchestration. Portkey partnered with specialist guardrail platforms, acknowledging that steering and evaluating model behavior required expertise beyond gateway operations.[4]

On February 19, 2026, Portkey announced it had raised $15 million in Series A funding, led by Elevation Capital with participation from Lightspeed. It described the gateway as a unified control plane that enforced policy, routed traffic, observed requests and tracked spend in real time. The company also announced free availability of its basic enterprise gateway and identified enterprise go-to-market expansion, governance for agentic AI and infrastructure for larger workloads as investment priorities.[5]

In that February announcement, Portkey reported more than 500 billion tokens and 120 million requests daily, more than 24,000 organizations, and support for more than 60 providers and 1,600 models. It also reported more than $500,000 in daily managed spend and more than $180 million in annualized spend under management. These spending figures describe AI expenditure managed through the platform; the organization count does not establish a paying-customer count.[5]

On March 24, 2026, Portkey announced it was open-sourcing its Production Gateway. The announcement, written by Swetha Sridhar, moved circuit breakers, semantic caching, budget limits, a model catalog, metadata governance and configuration management into the open-source offering from features previously requiring a SaaS subscription. It separately described an On-Prem Enterprise Gateway with additional enterprise controls.[6]

The expanded gateway addressed agent workflows with an MCP server registry for tracking and versioning servers, plus OAuth 2.1 authentication using PKCE. Circuit breakers could respond to P99 latency or error rates, use probe requests to test recovery, and fall back to cached responses, another provider or an error. Usage policies enforced request, token or cost limits at ingress, before calling the model. The March announcement reported one trillion tokens processed daily, a separately dated traffic figure from February’s report.[6]