Phoenix evaluates your model. We measure your product and your agent.
Different jobs, and we will say so plainly: Phoenix is deeper than us on OpenTelemetry-based LLM tracing, evals, datasets and prompt tooling, it is genuinely light to self-host, and it is free. smolanalytics is one MIT Go binary that does product analytics and computed agent observability from the same events, with numbers no model ever labels and no metered AI.
smolanalytics vs Arize Phoenix, which one do I want?
It depends on what you are debugging, and plenty of teams run both. Arize Phoenix is Arize AI's source-available AI observability and evaluation platform: OpenTelemetry and OpenInference-based tracing of every model and agent call, LLM-based and code-based evals, versioned datasets, experiments, and prompt management with a playground. It is deeper than smolanalytics at all of those, it self-hosts light, one container with SQLite by default and Postgres for production, or just pip install arize-phoenix, and Arize says self-hosting is free with no feature gates. Two things to know before you pick it: the license is the Elastic License 2.0, which is source-available rather than OSI-approved open source and does not allow offering Phoenix as a managed service, and the enterprise path is Arize AX, which Arize describes as the enterprise SaaS version of Phoenix; Arize raised a $70M Series C in February 2025, so the company behind it is well funded. If you need trace-level LLM debugging, evals, datasets or prompt versioning, use Phoenix. smolanalytics is a different shape: a single MIT Go binary that does web and product analytics (funnels, retention, paths, cohorts, breakdowns, heatmaps, surveys, feature flags and A/B) and computed agent observability off the same event stream, tool-call volume, error rate, latency p50/p90/p99, client split, an error taxonomy, and conversation health like turns, re-ask rate, abandonment and resolution. Those numbers are computed from your events, never labelled by a model, and a CI test pins /v1, MCP and the dashboard to the same answer. You ask it questions in plain English from your editor: one MCP connection with an org token operates every project, 79 tools, answered by your own coding agent on your own model, so the AI is never metered by us. Pricing is a 14-day full trial with no card, Pro $49/mo for 1M events, Scale $149/mo for 10M, $8 per extra million, self-host free forever.
what each one actually is
“AI Observability & Evaluation”: OpenTelemetry-based tracing, LLM evals, versioned datasets, experiments, and prompt management with a playground, under the Elastic License 2.0.
github.com/Arize-ai/phoenix, repo description, README and license
“Self-hosting on your own infrastructure or in your cloud account is free and fully permitted.” “There are no feature gates — Phoenix is a fully open-source platform.” The license is ELv2, so it is source-available, not OSI open source.
arize.com/docs/phoenix, self-hosting license page
smolanalytics is one MIT Go binary and one data file. Product analytics and agent observability come out of the same events, every number computed, none of them a model's opinion.
what we do instead
Arize Phoenix vs smolanalytics
| Arize Phoenix | smolanalytics | |
|---|---|---|
| The job it is built for | LLM traces, evals, experiments | product analytics + agent health |
| Trace-level LLM debugging | deep: OTel traces, replay in playground | no, we compute rollups |
| Evals, datasets, experiments | core product, LLM and code evals | not offered |
| Prompt management | versioning, tagging, playground | not offered |
| Funnels, retention, paths, cohorts | not its category | yes, same binary, same events |
| Agent numbers | traces plus model-scored evals | computed: p50/p90/p99, error taxonomy, re-ask, resolution |
| Self-host footprint | one container, SQLite or Postgres | one static binary, one file |
| Ask from your editor | Phoenix MCP server | MCP, one org token routes every project, 79 tools |
| Dashboard vs API drift | separate read paths | CI pins /v1 == MCP == dashboard |
| License | Elastic License 2.0, source-available | MIT, all of it |
when to use Phoenix (often alongside us)
- You need to debug a single run call by call. Phoenix gives you the OpenTelemetry trace of every step, the inputs and outputs, and a playground to replay traced LLM calls. We compute rollups over your events and do not store per-call LLM traces, so if the question is “what exactly happened inside that one run”, that is Phoenix, not us.
- You run evals or manage prompts. LLM-based and code-based evals, versioned datasets, experiments, and prompt management with version control and tagging are core Phoenix products. We ship none of them and are not trying to.
- Your stack is OpenTelemetry-first, or you want the notebook workflow. Phoenix instruments over OTel and OpenInference, starts with a pip install or a single container, and is free to self-host with no feature gates, with well-funded Arize behind it ($70M Series C, February 2025) and Arize AX as the enterprise path if you outgrow it. Running both is a normal setup: Phoenix for the model layer, smolanalytics for the product and the computed agent health beside it.
Add the other half in 30 seconds
# nothing to instrument, no OTel collector, no evals to configure docker run -p 8080:8080 -v $PWD/data:/data \ -e SMOLANALYTICS_DB=/data/smolanalytics.data \ ghcr.io/arjun0606/smolanalytics # dashboard + MCP at http://localhost:8080 # then, in your editor: "did signups who hit an agent error this week come back?"
Keep Phoenix for traces, evals and prompt experiments. Put this next to it for funnels, retention and the computed agent numbers, on your own box, answered by your own model.
questions
Is smolanalytics an Arize Phoenix alternative?
Only partly, and we would rather be straight about it. Phoenix is an AI observability and evaluation platform: OpenTelemetry-based traces, LLM and code evals, versioned datasets, experiments and prompt management, and it is deeper than us at every one of those. smolanalytics is one MIT Go binary that gives you web and product analytics (funnels, retention, paths, cohorts, heatmaps, surveys, flags and A/B) plus computed agent observability from the same events: tool-call volume, error rate, latency p50/p90/p99, client split, an error taxonomy, and conversation health like turns, re-ask rate, abandonment and resolution. If the tracing and evals are what you need, use Phoenix. If you want product usage and agent health in one place, use us. Many teams run both.
Can smolanalytics replace Phoenix for LLM tracing and evals?
No. We do not store per-call LLM traces, we do not ship evals, datasets, experiments or a prompt playground, and we are not planning to. Our agent observability is computed rollups over the events your app already sends, which answers “is the agent healthy and where is it failing”, not “show me every span of run 4f2a”. For that second question, Phoenix.
Is Arize Phoenix open source?
Source-available, under the Elastic License 2.0. Arize's own license page says self-hosting on your own infrastructure or cloud account is free and fully permitted with no feature gates, and there is nothing hidden about that. But ELv2 is not an OSI-approved open-source license, and its defining restriction is that you may not offer the software to others as a managed service. For most teams self-hosting for themselves, that changes nothing in practice. smolanalytics is MIT, all of it, with no equivalent restriction. The commercial sibling is Arize AX, which Arize describes as the enterprise SaaS version of Phoenix, and a hosted Phoenix Cloud also exists.
How are smolanalytics agent numbers different from Phoenix evals?
Ours are computed, not labelled. Phoenix's evals use LLMs, code checks or human labels to score output quality, which is genuinely useful and something we do not do. Our numbers, tool-call counts, error rate, latency percentiles, client split, the error taxonomy, turns, re-ask rate, abandonment and resolution, are all derived deterministically from your event stream, so nothing is a model's opinion and nothing changes when a model changes. A CI test asserts the /v1 API, the MCP answer and the dashboard return the same number for the same question, so the answer you get in your editor cannot drift from the chart.
What does smolanalytics cost, and can I self-host it?
14-day full trial, no card. Pro is $49/mo for 1M events, Scale is $149/mo for 10M, and extra usage is $8 per million. Self-hosting is free forever under MIT: one Go binary, one data file, no external services. There is no free cloud tier.
Get the verdict, not another dashboard.
Drop one snippet, ask in your editor, own your data. 14-day full trial, no credit card.