Helicone

Overview

Helicone began as a one-line proxy for logging OpenAI requests and grew into an open-source AI gateway and observability suite. Its product followed the operational needs of AI teams: first cost and request visibility, then routing, fallbacks, prompt management, evaluations, and release workflows. By March 2026, the company reported processing 14.2 trillion tokens for 16,000 organizations and tracking 33 million end users.^[1]

Mintlify acquired Helicone on March 3, 2026. The outcome was not a shutdown, but it was not an independent growth ending either: the founders joined Mintlify, Helicone entered maintenance mode, and customers were offered support migrating elsewhere.^[2] The strongest reading is that Helicone won adoption while the standalone observability category lost strategic altitude. Basic tracing became available from open-source competitors and model providers; the more valuable prize moved toward the knowledge infrastructure that agents use to act.

Founding Story

Helicone came out of Y Combinator's Winter 2023 batch in San Francisco. YC identifies Justin Torre, Barak Oshri, and Scott Nguyen as founders. Torre had worked in developer evangelism and teaching at Apple, Oshri had been a machine-learning engineer at Sisu Data and worked in Stanford's AI Lab, and Nguyen brought user-experience and finance experience.^[3] Cole Gottdank later appeared as a founder and co-signatory of the acquisition announcement. Public sources reviewed here do not explain the founder lineup change or how the original team met.

The founding wedge was deliberately small. Developers changed a base URL, routed model calls through Helicone, and received logs plus cost, user, model, and prompt analytics. Caching and intelligent retries addressed reliability and spend without forcing a team to replace its application stack.^[3] That low-friction entry mattered because AI application teams in 2023 were shipping before a stable tooling category existed.

The founders later wrote that YC batchmates asked for access, continued using the product, and referred friends. They described Helicone as an early mover before “LLM observability” had settled into a recognized market.^[1] The evidence handoff does not contain two complete, attributable founder quotations, so this report does not manufacture them. That source gap limits what can be said about the founders' private motivations and internal decisions.

Timeline

• Winter 2023: Helicone joined Y Combinator and launched around proxy-based LLM logging, analytics, caching, and retries.^[3]
• January 23, 2025: Helicone launched V2 on Hacker News as a workflow spanning “Log → Evaluate → Experiment → Review → Release.” It reported 2.1 billion requests and 2.6 trillion tokens processed.^[4]
• March 3, 2026: Mintlify announced its acquisition of Helicone. Justin Torre and Cole Gottdank joined Mintlify; Helicone moved into maintenance mode.^[1]

What They Built

Helicone sat between an AI application and model providers. A developer could route a request through its gateway and gain a record of what was sent, what came back, latency, cost, user context, and errors. That architecture made instrumentation an infrastructure choice rather than a recurring application-code project.

The product expanded in both directions. At the traffic layer it added routing, automatic provider fallback, retries, caching, and cost controls. At the development layer it added prompt versioning, evaluations, experiments, agent-session debugging, and review-to-release workflows. Its current open-source repository describes a Next.js web app, a Cloudflare Workers proxy, an Express/Tsoa collector, Supabase for database and authentication, ClickHouse for analytics, and object storage through Minio.^[5]

Scale forced architectural change. Helicone said synchronous log processing overloaded its database and pushed queries from milliseconds to minutes. The team rebuilt ingestion around Kafka and split storage by access pattern across Kafka, S3, and ClickHouse. It reported zero data loss and fast queries at billions of records after the change.^[4] This was more than backend housekeeping: a monitoring product that drops events or stalls under load destroys the trust it sells.