5 Arize Phoenix Alternatives for Permissive Self-Hosting in 2026

Phoenix is the default open-source pick for LLM tracing for good reasons. It starts in under a minute, it is built natively on OpenTelemetry, and its RAG eval is widely called the best in the category. I recommend it often. But there is one line in its marketing that does not hold up, and if you are here, it is probably why.

Arize markets Phoenix as “fully open source, no feature gates.” The main server repo is actually Elastic License 2.0 - source-available, not OSI-approved open source. Only the client and eval subpackages - the ones you might pull into your own code - are Apache-2.0. ELv2 forbids offering Phoenix as a hosted or managed service to third parties. For internal use that restriction rarely comes up, and the features genuinely are not gated. But it is not the freedom that MIT or Apache-2.0 gives you, and “fully open source” overstates it. If you need a genuinely permissive self-host, here are the alternatives, matched to why you left.

The short version

Tool	Best for	License	Self-host	Starting price
Arize Phoenix	Best RAG eval, OTel-native	ELv2 server	Free, runs locally	Free
Opik	Most permissive OSS, cheapest cloud	Apache-2.0	Free, no gates	Free / $19/mo
Langfuse	MIT self-host, cheapest at scale	MIT	Free, full-featured	Free / $29/mo
DeepEval	Research-backed eval metrics	Apache-2.0 (OSS)	Free framework	Free / $200/mo
LangSmith	Turnkey, no OSS debate	Closed	Enterprise only	$39/seat/mo

If you left because you want a truly permissive license: Opik

This is the direct swap, so it goes first. Phoenix’s server is ELv2. Opik is Comet’s Apache-2.0 platform, and its OSS build ships the full feature set self-hosted, with unlimited spans, members and retention, and no feature gates. Apache-2.0 is a genuinely permissive OSI license - none of the “you may not offer it as a service” restriction that ELv2 carries. It is also the fastest-growing project of its peers, per the vendor, at roughly 20.8k GitHub stars. And if you want managed instead, the Pro cloud is $19/mo for 100k spans, the cheapest paid tier of the major eval platforms.

The gotcha is per-seat pricing at scale. The $19 headline is the small-team configuration - the recurring complaint is that seat costs add up as your team grows. A few users also report UI slowdown on very large projects. And Opik treats OpenTelemetry as one ingestion path rather than the native architecture, so if strict OTel-native was part of why you liked Phoenix, Opik is a small step back there. On the license question, though, it is the cleanest answer in the category.

If you want MIT and the cheapest cost at scale: Langfuse

If Apache-2.0 is good but you specifically want MIT, or your trace volume is high enough that cost dominates, Langfuse is the move. It is MIT-licensed, only three features are enterprise-gated in the self-host build, and at 1M events a month it runs about $101/mo managed - roughly 25x cheaper than LangSmith for comparable volume, per widely-cited comparisons. Like Phoenix it runs as an OpenTelemetry backend, mapping the GenAI semantic conventions.

The catch is operational. Langfuse v3 needs Postgres plus ClickHouse, Redis and S3-compatible storage - four services, versus Phoenix’s famous run-it-on-a-laptop start. GitHub has a string of migration threads where that architecture trips people up. gRPC OTLP is not supported yet, HTTP only. And per the vendor, Langfuse became a ClickHouse subsidiary in January 2026, with the MIT license preserved - worth filing away for a multi-year bet. If you cannot run four services, the $29 Core cloud sidesteps it.

If the RAG eval was the reason you were on Phoenix: DeepEval

A lot of people are on Phoenix for the eval library, not the tracing. If retrieval quality is your bottleneck and you want an Apache-2.0 framework you fully own, DeepEval is the closest match. Its open-source framework is Apache-2.0 and ships 50-plus research-backed metrics including G-Eval, with dedicated RAG scorers like faithfulness and answer relevancy - and it drops into a Python test suite like pytest, so it runs in CI. It is an independent YC startup, not owned by a model vendor, which some buyers weigh.

The gotcha is the cost of judging. Nearly all of DeepEval’s metrics are LLM-as-judge, so every test case fires another inference. Large suites rack up real API bills and runs take minutes, not seconds, so CI needs explicit timeouts. The Confident AI cloud on top also has a steep cliff - Starter is $200/mo and the next rung, Team, is $2,000/mo, a straight 10x jump per the vendor’s pricing. Live in the free framework as long as you can. On eval quality specifically, Phoenix versus DeepEval is a genuine toss-up.

If you are done with the OSS license debate entirely: LangSmith

Maybe the honest answer is you do not want to read another license. If you have budget and want turnkey, LangSmith has no OSS asterisks to argue about - it is simply a managed product. The tracing is the deepest available for LangChain and LangGraph, and Align Evals, its workflow for calibrating an LLM judge against human scores, is genuinely useful.

But you trade the license headache for a pricing one. LangSmith is fully closed source, self-hosting is Enterprise-only, and the bill runs roughly $2,514/mo at 1M base traces on one seat - about 25x Langfuse, per widely-cited comparisons. Backed by a company valued at $1.25B, per press reports, so survival is not the question. Value-for-money is. This is the pick only if you were never really attached to open source and Phoenix’s license was an annoyance rather than a principle.

So which one?

• You want a genuinely permissive license with no service restriction - Opik, Apache-2.0 with no gates, or Langfuse if you specifically want MIT.
• Trace volume is high and cost dominates - Langfuse, self-hosted or the $29 Core cloud.
• RAG eval quality is the whole point - DeepEval for an Apache-2.0 framework you own, though budget for the LLM-judge API bill.
• You are tired of arguing licenses and have budget - LangSmith, turnkey and closed, if the trace bill does not scare you.
• The ELv2 restriction never actually applies to you - honestly, stay on Phoenix. For internal use it behaves like open source, the RAG eval is the best around, and the sub-minute start is unbeaten. The license only bites if you meant to resell it as a service.

Every price and license above was read from each vendor’s own materials on 26 July 2026. This category ships breaking changes monthly, so we re-verify every 30 days. The Phoenix license nuance is real, but for most teams running it internally, it is a footnote - not a reason to migrate.