Helicone: proxy-based LLM logging and gateway, now in maintenance mode

Helicone logs LLM traffic by sitting in the request path: point your SDK at Helicone instead of the provider and every call is recorded with cost, latency and tokens. On 3 March 2026 Helicone announced it was joining Mintlify, and the product moved to maintenance mode. Read this page as a guide to running what you have and planning what comes next.

Helicone is an observability platform with a gateway built in. There are three ways in:

• The AI Gateway. Set your OpenAI SDK's base URL to https://ai-gateway.helicone.ai and use a Helicone API key. You can call 100+ models with one key: either Helicone holds the provider keys and you buy credits at 0 percent markup, or you bring your own provider keys, which are always tried first.
• Legacy provider proxies. Keep your existing SDK, change api.openai.com to oai.helicone.ai (or anthropic.helicone.ai) and add a Helicone-Auth header. Helicone marks these as maintained but no longer actively developed.
• Async logging. Helicone's async loggers, built on OpenLLMetry, send logs after the call instead of proxying it, so a Helicone outage cannot take your app down. You lose the gateway features: caching, retries and rate limits.

Behavior is controlled with HTTP headers on each request: sessions, user IDs, custom properties, caching, retries, rate limits, omitting content, and security checks. In the stack it straddles the gateway and observability layers.

The status matters more than any feature. The founders' announcement says Helicone's services "will remain live for the foreseeable future in maintenance mode", with security updates, new models and bug and performance fixes still shipping. The repository is Apache-2.0 and still receives security fixes, but the last tagged server release is from August 2025 and the helper packages were last published in November 2025. No roadmap or end date has been announced.

Keep it, or pick it, when:

• You already run Helicone and it does the job. Nothing forces a move today, and the cloud service is live. Use the time to plan a calm migration rather than a rushed one.
• You need request logs from several services this week and cannot touch their code beyond a base URL and a header.
• You want edge caching of identical requests during development, where repeated test prompts cost real money.

Choose something else when:

• You are picking a platform for the next few years. New models keep arriving, but a product in maintenance mode is unlikely to follow new agent frameworks or evaluation methods, and may lag on new provider APIs.
• You need agent tracing. Helicone is request-centric: sessions are assembled from headers on each logged call (model requests, plus tool and vector-database calls you log with its manual logger). Span-based tools capture your own functions and control flow as nested spans automatically.
• You need evaluations. Helicone stores scores you compute elsewhere; its docs state it is not an evaluation framework.
• You want to self-host at scale. The Docker image is easy, but the README routes production Helm requests to Helicone's enterprise contact, the public chart repository has not changed since September 2025, and the self-hosting docs changed as the proxy moved out of the API server.

For new work, the usual replacements are LiteLLM for the gateway (keys, budgets, fallbacks, caching) and Langfuse, Arize Phoenix or Laminar for traces and evals. Instrumenting with OpenLLMetry or the OpenTelemetry GenAI conventions keeps you free to switch backends later.

Requests go through the edge. The gateway translates OpenAI-format requests to each provider, forwards them, returns the response and logs the call. Helicone documents the logging delay as negligible, but the gateway is in your critical path: if it is down or slow, so are you.

Routing is in the model string. A plain model name routes to the cheapest provider that serves it and fails over on 429, 401, 400 context-length, 408 and 5xx errors. A suffix pins a provider (gpt-4o-mini/openai in Helicone's own example), a deployment ID pins one of your own deployments, a comma-separated list is an explicit fallback chain, and a !provider prefix excludes a provider.

Headers do the rest:

Responses come back with Helicone-Cache (HIT or MISS) and a helicone-id you use to attach scores later. Prompts can live in Helicone and be referenced by prompt_id and inputs through the gateway.