MLflow AI Gateway vs Helicone

Why

These two reached the same place from opposite directions. MLflow started in 2018 as experiment tracking and added a gateway to the server that already held your traces, so a gateway request becomes an MLflow trace with no extra instrumentation. Helicone started in 2023 as LLM request observability and added a gateway in front of the logging it already did, so the request is the primary object and the analytics hang off it. The economics are close enough to ignore: both are Apache-2.0, both charge zero token markup, both charge $0 per seat, and both record deployment as hosted or self-hosted. Helicone's paid tiers exist for the platform rather than the routing — Hobby free with one seat, 10,000 requests a month, 1 GB of storage and 7-day retention, then Pro at $79 a month and Team at $799 — while MLflow has no paid tier at all.

The deciding question turns out not to be lineage but momentum. Mintlify announced its acquisition of Helicone on 2026-03-03 and put the product into maintenance mode: security updates, bug fixes and new models keep shipping, Helicone says the services remain live for the foreseeable future, and Mintlify says it will work closely with every customer to support a smooth migration to another platform. The dates in the record are consistent with that. The latest tagged GitHub release is 2025-08-21, the hosted changelog stops at 26 November 2025, and the standalone Rust gateway repository was last pushed 2025-11-21. MLflow's latest release in the same record is dated 2026-08-26. This is not a reason to write Helicone off — it is a reason to be honest that you would be adopting a product whose owner has offered to help you leave it, and the signup flow still advertises a free trial, so the status is easy to miss.

If you self-host, provider coverage inverts against the star counts. Helicone's hosted registry returned 111 models across 21 provider endpoints, and its docs describe the gateway as 100+ providers; the self-hosted build supports OpenAI and Anthropic only, with Vertex AI, AWS Bedrock and Azure OpenAI explicitly not supported, and port 8585 carries no proxy authentication so firewall restriction is required. MLflow enumerates 14 providers in its docs tables — including Azure OpenAI, AWS Bedrock and Vertex AI — and self-hosting is the first-class path, by pip or an official OCI Helm chart, with the same 14 available either way. So Helicone's 6,109 stars against MLflow's 27,777 tell you very little here, not least because those 27,777 are for all of mlflow/mlflow and not a measure of gateway adoption. MLflow publishes no model total, and points at the provider dropdown instead of a number.

What is left is a genuine trade between logging depth and enforcement. Helicone logs prompt and completion bodies by default, attributes cost per user through a Helicone-User-Id header and per property through Custom Properties, and ships sessions, HQL, alerts and a public model registry with per-endpoint pricing; it records config-as-code as supported, exact-match caching, SOC 2 Type II, a HIPAA BAA, GDPR and a choice of US or EU region. MLflow is quieter by default: logging is opt-in per endpoint and off until you enable usage tracking, at which point full payloads are recorded as traces in your own database, and cost attribution is per endpoint, provider and model rather than per user. Enforcement runs the other way. Helicone records guardrails and PII redaction as absent, its one request-path control being an opt-in Helicone-Moderations-Enabled header that a client can simply omit, while MLflow ships LLM-judge Safety, PII and Custom guardrails with Block or Sanitize actions — subject to the streaming caveat below. Neither is the choice if you want declarative endpoint management on the MLflow side, which records config-as-code as unsupported.