What it is
Opik, from the ML experiment-tracking company Comet, traces LLM calls and agent steps, runs evaluations with LLM-as-judge metrics, manages prompts and datasets, and monitors production with dashboards and online evaluation rules. A pytest integration runs evaluations in CI, an Agent Optimizer SDK tunes prompts automatically, and an MCP server lets Claude Code or Cursor read traces from chat.
The whole platform, server and web app included, is Apache 2.0, and Comet's pricing page describes the self-hosted build as the same codebase as the hosted one. Self-hosting runs on Docker Compose for local use or a Helm chart for production, on a stack of ClickHouse, MySQL, Redis, ZooKeeper and MinIO.
Hosted plans start with a free tier of 25,000 spans a month and 60-day retention. Pro costs $19 a month for 100,000 spans, then $5 per 100,000; retention up to 400 days is an add-on priced at $29 per 100,000 spans.
Our review
Opik's case rests on licence and price. Every part of it, server included, is Apache 2.0 with no enterprise folder, and the Pro cloud plan at $19 a month for 100,000 spans undercuts Langfuse Core at $29. For a team whose lawyers want a plain permissive licence on the thing they host, that settles a lot.
The feature list tracks Langfuse closely: traces, datasets, experiments, LLM-as-judge metrics for hallucination and moderation, online evaluation and prompt management. Its additions are the Agent Optimizer, a pytest integration that makes evaluations behave like unit tests, and AI Guardrails, which the pricing page lists only for the self-hosted edition. Integrations run to more than 60 frameworks and providers, plus OpenTelemetry for Java, Ruby and .NET. The README's claim of 40 million traces a day is a vendor figure with no published benchmark behind it.
The self-hosted edition has a gap that the licence talk hides. Comet's own docs say it comes "without user management features", so there are no logins or roles: anyone who can reach port 5173 sees every trace, and traces usually hold prompts and customer text. It also sends a daily anonymous usage report to Comet by default (trace, experiment and dataset counts plus a server ID) until OPIK_USAGE_REPORT_ENABLED is set to false. Hosted SSO, RBAC and the SOC 2, ISO 27001 and HIPAA line all sit on the Enterprise plan.
Release cadence is near-daily; 2.2.119 shipped on 23 October 2026, and the docs warn that newer SDKs may need a newer server, so a self-hosted install has to keep pace. With about 22,800 GitHub stars to Langfuse's 36,200, Opik has fewer community answers to lean on. The pricing page still contradicts itself on seats, promising unlimited members in the FAQ while capping Free at 10 and Pro at 50 on the plan cards.
Where it earns its place
An ML engineer at a fintech that cannot send prompts to a US SaaS
Trace and evaluate an internal RAG assistant without data leaving the VPC
Deploy the Helm chart into the existing Kubernetes cluster and point the Python SDK at it with opik configure. No licence fee, but the team runs ClickHouse and MySQL itself and must put SSO or a VPN in front, because the open-source edition has no user accounts.
A three-person startup shipping a customer-facing agent
Get production traces and online LLM-judge scores without running infrastructure
Pro at $19 a month covers 100,000 spans; at 2 million spans a month the bill is about $114 at $5 per extra 100,000. Data is stored in the US and kept 60 days unless the retention add-on is bought.
A platform team already using Comet for model training
Keep LLM evaluation next to existing experiment tracking
Opik runs in the same Comet workspace, and the pytest integration fails a CI build when a hallucination or relevance metric drops. The MLOps add-on and OpikAssist are cloud-only, so mixed self-hosted setups lose them.
Before you commit
- Setup
- pip install opik and add the @track decorator for the hosted cloud. Self-hosting is git clone plus ./opik.sh on Docker Compose, which the docs call not production-ready; production uses the Helm chart. Comet publishes no CPU or memory sizing.
- Integrations
- Python and TypeScript SDKs, a REST API, OpenTelemetry ingest, and documented integrations for more than 60 frameworks and providers including LangChain, LangGraph, LlamaIndex, OpenAI Agents, Vercel AI SDK, LiteLLM, Dify and n8n. An MCP server connects Claude Code, Cursor and Codex.
- Data and privacy
- Self-hosted traces stay in your ClickHouse, but the server sends a daily anonymous usage report to Comet unless OPIK_USAGE_REPORT_ENABLED=false, and there is no built-in login. A medium-severity SSRF in the alert webhook test endpoint was published in September 2026. Cloud is US-hosted; SOC 2, ISO 27001 and HIPAA are listed on Enterprise only.
- Cost at scale
- Self-hosted is free apart from the cluster. Cloud Pro is $19 plus $5 per 100,000 spans over 100,000, so 10 million spans a month is about $514; extending retention to 400 days costs $29 per 100,000 spans. SSO, RBAC and SLAs need a custom Enterprise contract.
- Limits
- 181 open issues and pull requests, many on LLM-judge score parsing and dashboard filters. Near-daily releases and SDK-server version coupling mean self-hosted installs need regular upgrades. Smaller community than Langfuse.
What's good
- Apache 2.0 throughout, including the server
- Pro cloud at $19 a month undercuts Langfuse Core ($29)
- Pytest integration, prompt optimiser and self-hosted guardrails
- OpenTelemetry ingest for Java, Ruby and .NET
What's not
- Self-hosted edition has no logins or roles
- Anonymous usage report on by default
- SSO, RBAC and compliance only on Enterprise
- Pricing page contradicts itself on seat limits
Best for
Teams that want a permissively licensed, self-hostable tracing and evaluation platform, or a low-cost hosted one.
Alternatives
Opik: common questions
- Is Opik worth it?
- Opik gives away the whole server under Apache 2.0 and sells the cheapest hosted plan, but the self-hosted build has no user accounts and phones home by default. We rate it 4.1 out of 5.
- How much does Opik cost?
- Free and open source under Apache 2.0. Cloud: Free (25k spans/month, 60-day retention); Pro $19/month (100k spans, then $5 per 100k; 400-day retention add-on $29 per 100k spans); Enterprise custom.
- What are the downsides of Opik?
- Self-hosted edition has no logins or roles; anonymous usage report on by default; sSO, RBAC and compliance only on Enterprise; pricing page contradicts itself on seat limits.
- Who is Opik best for?
- Teams that want a permissively licensed, self-hostable tracing and evaluation platform, or a low-cost hosted one.
- How much does Opik cost for a team or in production?
- Self-hosted is free apart from the cluster. Cloud Pro is $19 plus $5 per 100,000 spans over 100,000, so 10 million spans a month is about $514; extending retention to 400 days costs $29 per 100,000 spans. SSO, RBAC and SLAs need a custom Enterprise contract.
- Is Opik safe for company data?
- Self-hosted traces stay in your ClickHouse, but the server sends a daily anonymous usage report to Comet unless OPIK_USAGE_REPORT_ENABLED=false, and there is no built-in login. A medium-severity SSRF in the alert webhook test endpoint was published in September 2026. Cloud is US-hosted; SOC 2, ISO 27001 and HIPAA are listed on Enterprise only.
- What are the best Opik alternatives?
- The closest alternatives we have reviewed are Langfuse, Arize Phoenix and DeepEval.
Answers checked Oct 23, 2026.
Make Opik? Show this rating on your site
