MLflow AI Gateway vs Braintrust Gateway

Why

Both of these are a routing surface attached to something larger, and in both cases the something larger is the reason you would adopt it. MLflow's gateway is a feature of the Tracking Server rather than a deployable of its own: as of MLflow 3.0 the standalone deployment server application and the start-server CLI command were removed, so mlflow server is the only entry point. Braintrust's Gateway sits inside its eval and observability platform, with the standalone proxy MIT-licensed on GitHub and the platform itself recorded as open core. The consequence is that neither should be evaluated as a proxy in isolation. You are choosing the platform, and inheriting its gateway.

The licence and the bill diverge from there. MLflow is Apache-2.0 under Linux Foundation governance with no pricing page and no paid gateway tier at all: zero token markup, zero credit fee, zero per seat, and your only cost is the Tracking Server plus a SQL backend and, in production, object storage. Braintrust also charges $0 per seat and lists the hosted Gateway as free during public preview, but pricing is to be announced before general availability, so the number you would plan around does not exist yet. What is published is the platform around it: Starter at $0 with $10 of model credits, 1 GB of processed data, 10k scores and 14-day retention, then Pro at $249 a month, which includes $249 of credits dropping to $100 a month after 2026-09-01, with processed-data overages at $4 per GB on Starter and $3 per GB on Pro and scores at $2.50 and $1.50 per thousand.

On enforcement the two records point in opposite directions, and this is the sharpest functional split. MLflow records guardrails as present: Safety, PII and Custom checks configured per endpoint with Block or Sanitize actions, evaluated by an LLM judge that is itself another gateway endpoint, so the decision happens inside your own server while the judging inference goes to whatever provider backs the judge. Braintrust records guardrails as absent, and deliberately so — scorers, spans and online scoring run after the fact, with no request-path enforcement, and its PII handling is a global masking function that redacts before logging rather than before the provider sees the prompt. If your requirement is that something never leaves the boundary, only MLflow attempts it. The caveat is load-bearing: post-LLM guardrails are not triggered for streaming requests, so a Safety guardrail on its default stage stops enforcing the moment a client sets stream true.

Everything else is the asymmetry you would expect between a foundation project and a funded company. Braintrust publishes SOC 2 Type II, a HIPAA BAA and EU residency on a dedicated EU host in eu-west-1, and raised an $80M Series B led by ICONIQ in February 2026; MLflow's certification fields are all recorded not applicable, because there is no vendor to attest anything. Self-hosting inverts it: MLflow installs by pip or an official OCI Helm chart with TLS, ingress, Prometheus metrics, NetworkPolicy and RBAC from 3.13.0, while Braintrust self-hosting is a managed control plane plus a customer-hosted data plane and is effectively Enterprise-only. Read the headline counts carefully. MLflow's 27,777 stars are for all of mlflow/mlflow and are not a measure of gateway adoption, against 409 on Braintrust's proxy repo; MLflow enumerates 14 providers in its docs tables while other vendor pages say 100+ and 50+, and publishes no model total, against Braintrust's 18 providers and over 100 models. MLflow's 28.6 ms of P50 overhead and 598 requests per second are its own published figures from its own LiteLLM comparison, measured with a 50 ms simulated provider delay, 4 workers and 50 concurrent users; Braintrust publishes no overhead number at all.