Langfuse alternatives for understanding users, not just traces

Stay on Langfuse for tracing, evals, and prompt management: it is MIT-licensed, free to self-host, and the better choice than every managed tool here if open-source tracing is your requirement. What it does not do out of the box is answer product questions such as “what are users trying to do” and “where do they give up”, so most teams keep Langfuse and add one tool for that layer, reading the same OpenTelemetry data. Replace Langfuse only if you want a single vendor for both jobs. We make Agnost AI, so it is listed first; every tool gets the same structure and its limitations come from its own documentation.

Langfuse vs the alternatives for understanding users

The columns are the three questions Langfuse leaves to you: what users wanted, whether they were frustrated, and whether they got what they came for.

Where Langfuse is stronger

Langfuse is the reference open-source platform for LLM tracing, and none of the tools below replaces all of it.

• Open source that is actually complete. The MIT version includes tracing, sessions, users, OpenTelemetry, prompts, datasets, LLM-as-judge, annotation queues, custom dashboards, enterprise SSO, and org-level RBAC. Enterprise-only features are mostly governance: server-side data masking, project-level RBAC, retention management, SCIM, and audit logs.
• Free self-hosting. Docker Compose, a VM, or Kubernetes with Helm, plus Terraform for AWS, Azure, and GCP.
• Prompt management, datasets, and experiments in the same product as tracing.
• A per-user view. Langfuse lists users and segments them by token usage, trace count, and feedback, with all of a user’s traces in one place. For “which users cost the most” or “which users left a thumbs-down”, it is already enough.
• Low entry price. Cloud Hobby is free and Core is $29 a month with unlimited users.

The gap this page is about. Conversation themes, intent labels, and outcomes are not built in. Langfuse’s own guide for product teams says to “run an intent classification pipeline that writes labels back as categorical scores”, treats resolution as a score filled in by reviewers or feedback, and mentions sentiment only as an example criterion for a custom judge. Live evaluators are recommended at the observation level, trace-level evaluators are deprecated, and session-level live evaluation is not described, which makes whole-conversation judgments harder to run continuously. You can build all of this on Langfuse; the tools below ship some or all of it.

Agnost AI

What it is. Conversation analytics for production AI agents. Agnost AI reads real chat and voice conversations to show what users wanted, where they got stuck, and where the agent broke its own rules, then turns recurring failures into pull requests your team reviews.

Best for. Teams that already trace with Langfuse and want the product layer on top without building an intent pipeline: what users ask for, where conversations go wrong, and what to fix first.

Key features.

• Intents: plain-language definitions of what users want. Agnost classifies conversations against them, tracks volume over time, links to the evidence, and suggests new intents from your product description and observed traffic. This is the classification Langfuse’s guide asks you to build.
• Violations: rules written in natural language (“confirm identity before sharing account details”) checked against every analyzed conversation, with the supporting messages attached. Frustration patterns such as rephrasing loops or requests for a human are tracked the same way, as intents or rules you define.
• Conversation timelines that keep turns, tool calls, errors, and audio together, plus tool-call analytics for success rate, error distribution, and latency.
• Alerts to Slack or email on an intent or violation threshold, a Daily Summary, and a summary when a new user intent appears.
• Auto Improve: investigates a recurring finding, prepares a change, tests it against past runs, opens a pull request, and tracks the outcome after deploy.
• A hosted OpenTelemetry endpoint that can run next to your Langfuse exporter, plus Python, TypeScript, and Go SDKs, voice platforms (LiveKit, Vapi, Retell, Telnyx, Trellus), and the Agnost MCP server for Claude, Codex, and Cursor.

Limitations. Not open source; self-hosting is an Enterprise option, so it does not satisfy a strict open-source requirement. There is no built-in frustration score or conversation task-success metric: you define success with a rule or the success event field. No retention or cohort analytics. No built-in PII redaction before ingestion, so you redact in your app (data governance). It does not replace Langfuse’s prompt management or experiments.

Pricing. Free (1,000 events a month, 7-day retention), Starter $49/month (10,000 events, 30-day retention), Pro $499/month (1,000,000 events, 90-day retention), Enterprise custom with self-hosted VPC.

Amplitude Agent Analytics

What it is. Amplitude’s product analytics plus Agent Analytics, which observes what users ask an AI agent, evaluates where it fails, and ties traces and evals to conversion, retention, and revenue.

Best for. Teams that already standardize on Amplitude and want “what did users want and did they get it” inside the funnels, cohorts, and retention charts they use today.

Key features. Seven built-in session signals: Task Completed, Response Quality, Session Safety, User Intent, User Friction, Has Negative Feedback, and Data Quality Issues. Topic clustering across sessions on every plan. Custom LLM and code-based evaluators that run on all sessions (bring your own LLM key). OpenTelemetry GenAI spans flow in through a bridge, or use the Python and Node SDKs.

Limitations. Conversation text is kept 14 days on Free and Plus (30 days on Growth and Enterprise). Agent sessions are capped at 5,000 a month on Free and Plus. Sentiment isn’t a named signal, and some sections are still labeled beta. Not open source (its SDKs are MIT-licensed), and it runs in Amplitude’s AWS data centers in the US or EU rather than on your infrastructure.

Pricing. Free (2M events and 5,000 agent sessions a month). Plus starts at $0 with the same agent-session allowance. Growth and Enterprise are custom. Agent Analytics is billed on agent sessions, not events.

PostHog

What it is. PostHog’s AI Observability (formerly LLM analytics) captures every LLM call as a trace built on PostHog events, so each trace carries the user, their session replay, and any exceptions.

Best for. Teams that want user-level answers and an open-source core. It is the closest open-source option to a full user analytics stack: funnels, retention, and session replay sit on the same data as the AI traces.

Key features. Sessions via $ai_session_id; automatic clusters of similar traces with AI-generated titles; positive, neutral, and negative labels on user messages from a local model with no LLM cost; LLM-as-judge templates including Task outcome and User correction, run per generation, trace, or session; funnels and retention on the same data; OpenTelemetry support.

Limitations. Full prompt and response text is deleted after 30 days. Clustering needs at least 1,000 traces in the past week. LLM-judge evals need your own provider key, and trace-level user feedback is in beta. The core is MIT outside the ee/ directory, and self-hosting is free but officially unsupported. Prompt management and eval experiments are not its focus, so many teams keep Langfuse next to it.

Pricing. First 100,000 AI events a month free, then from $0.00035 per event, falling with volume. No per-seat charges. Each evaluation run counts as one event.

Braintrust

What it is. An evaluation-first platform where production traces become datasets and evals, with Topics for automatic classification of production traffic.

Best for. Teams that would replace Langfuse with a managed eval platform and want intent, sentiment, and issue clustering from the same vendor.

Key features. Topics classifies logs with built-in Task (user intent), Sentiment, and Issues facets plus custom facets, clusters similar requests into themes, and writes labels back to logs. Online scoring rules run scorers on production traces. Thumbs feedback attaches to spans. OpenTelemetry ingest.

Limitations. Topics needs at least 100 summaries before it clusters and regenerates daily, not in real time, and draws down monthly model credits. There is no dedicated per-user view (you filter or group logs by a metadata field such as metadata.user_id) and no built-in task-completion scorer; Braintrust’s docs suggest writing a trace-level judge for workflow completion. The platform is not open source (the SDK is Apache-2.0), and self-hosting is Enterprise-only.

Pricing. Starter free (1 GB processed data, 10k scores, 14-day retention, unlimited users). Pro $249/month (5 GB, 50k scores, 30-day retention). Billed on processed data and scores, not seats.

LangSmith

What it is. LangChain’s agent observability, evaluation, and deployment platform.

Best for. Teams building on LangChain or LangGraph who would consolidate tracing, evals, and conversation clustering in one vendor.

Key features. Threads group runs into conversations. Insights automatically clusters production traces, including full conversations, into usage patterns and failure modes. Thread-level online evaluators can judge intent, outcome, and trajectory. OpenTelemetry in both directions.

Limitations. Insights is on Plus and Enterprise, samples up to 1,000 traces per report, and runs as reports rather than continuously. Frustration and task outcome need custom attributes or evaluators. There is no dedicated user view; per-user analysis depends on storing user IDs as metadata. The platform is not open source, and self-hosting is an Enterprise add-on.

Pricing. Developer $0 (1 seat, 5,000 base traces a month). Plus $39 per seat per month (10,000 base traces), then $5 per 1,000 additional base traces (14-day retention), or $7.50 per 1,000 extended traces (180-day retention).

Arize Phoenix

What it is. Arize’s open-source AI observability platform for experimentation, evaluation, and troubleshooting, built on OpenTelemetry and OpenInference.

Best for. Teams that must stay self-hosted, including air-gapped or regulated environments, and want ready-made conversation judges instead of writing their own.

Key features. Sessions track a whole conversation in one thread with a chat-style view. The pre-built User Friction evaluator detects corrections, retries after an unsuccessful response, frustration, and challenges to unexplained actions, at span, trace, or session level. A Completeness evaluator checks whether a conversation covered every request. An on-demand error-analysis skill run through MCP groups observed problems into labeled annotations with counts. Self-hosting is free with no feature gates.

Limitations. Licensed under the Elastic License 2.0, which is source-available rather than OSI-approved, though Arize calls Phoenix open source. Friction and completeness are evaluators you run and log, not an always-on classification layer; Phoenix docs describe exporting traces, running judges, and logging results back. Automatic grouping of recurring failures (Signal) and stakeholder dashboards are Arize AX features. Per-user analytics are not documented. Phoenix does not evaluate incoming production traces automatically: its server evals run on dataset experiments, and online evals are an Arize AX feature.

Pricing. Phoenix is free to use and self-host. Arize AX Free is $0 (25k spans a month, 15-day retention), AX Pro is $50/month (50k spans, 30-day retention), and AX Enterprise is custom, with a self-hosted option.

Opik

What it is. Comet’s open-source observability and evals platform for agents, covering traces, threads, LLM-as-judge metrics, and online evaluation of production traffic.

Best for. Teams that want an Apache-2.0 platform with conversation-level judges they can switch on as rules, and a low-cost cloud option.

Key features. Threads group traces by thread_id. Conversation metrics include UserFrustrationMetric (“how frustrated the user became over the interaction”) and SessionCompletenessQuality (“whether the user’s top-level goals were met”). Online evaluation rules run at trace, thread, or span scope with sampling and filters; thread rules wait for a 15-minute cooldown by default. OpenTelemetry endpoint over HTTP.

Limitations. Topic or intent clustering of conversations is not documented, so “what users wanted” still needs your own classifier. Rules only apply to data logged after you create them. There is no per-user view: dashboards group trace metrics by tags, a metadata key, name, or error. The full platform, including online evaluation, is open source; Enterprise adds RBAC, SSO, and support SLAs.

Pricing. Open source is free with unlimited spans and retention. Free Cloud covers 25k spans a month for up to 10 members. Pro Cloud is $19/month for 100k spans, then $5 per 100k spans.

How to choose

If your team…	Start with
Keeps Langfuse for tracing and wants intents, rule violations, and fixes as PRs on top	Agnost AI
Already runs Amplitude and wants intent, friction, and task-completion signals in its funnels	Amplitude Agent Analytics
Wants user analytics plus an open-source core and session replay	PostHog
Would move to a managed eval platform with topic clustering	Braintrust
Builds on LangChain or LangGraph and wants one vendor for traces and clusters	LangSmith
Must stay self-hosted and wants a ready-made friction evaluator	Arize Phoenix
Must stay Apache-2.0 and wants frustration and session-completeness judges as rules	Opik
Only needs per-user usage and feedback	Stay on Langfuse; its user view covers it

If staying open source is non-negotiable, Phoenix, Opik, and PostHog are the closest options, but none of them gives you intent clustering and outcome tracking fully out of the box: you still configure judges or rules. If it is negotiable, the managed tools above ship more of that layer.

Using Langfuse with Agnost AI

Langfuse’s Python and TypeScript SDKs are a thin layer on the official OpenTelemetry client, and Langfuse also accepts OTLP at /api/public/otel. Agnost AI runs a hosted OTLP collector, so if your spans pass through an OpenTelemetry tracer provider, you can append Agnost AI as an additional span processor and both tools receive the same spans (OpenTelemetry docs):

import os
from opentelemetry import trace
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.http.trace_exporter import OTLPSpanExporter

trace.get_tracer_provider().add_span_processor(
    BatchSpanProcessor(
        OTLPSpanExporter(
            endpoint="https://otel.agnost.ai/v1/traces",
            headers={"X-Agnost-Org-ID": os.environ["AGNOST_ORG_ID"]},
        )
    )
)

Set a stable conversation ID on your spans (agnost.session_id or session.id); without one, Agnost AI treats each trace as its own conversation. Langfuse keeps your traces, prompts, datasets, and evals; Agnost AI classifies the same conversations against your intents and rules, alerts when a rate moves, and opens pull requests for recurring failures. To start, follow the quickstart.

Related reading

• Agnost AI vs Langfuse
• Best conversation analytics tools for AI agents (2026), the full ten-tool comparison
• LangSmith vs Langfuse vs Braintrust vs Phoenix vs Agnost, for tracing and eval workflows
• Braintrust alternatives for conversation analytics
• Intent analysis tools for AI conversations
• Self-hosted agent observability checklist

Frequently asked questions

Does Langfuse have user analytics?

Partly. Langfuse has a user list and a per-user view that segments by token usage, number of traces, and feedback, plus sessions and custom dashboards. It does not ship built-in intent or topic grouping, sentiment or frustration detection, or conversation outcomes, so questions about what users wanted and where they gave up need extra work.

Can Langfuse detect user intent or frustration?

Not out of the box. Langfuse's own guide for product teams suggests running an intent classification pipeline that writes labels back as categorical scores, and sentiment appears only as an example criterion for a custom LLM judge. You can build both, but you write and maintain the classifiers yourself.

Should I replace Langfuse or add another tool?

Most teams should add one. Langfuse is strong at open-source tracing, evals, and prompt management, and a conversation analytics tool can read the same OpenTelemetry data. Replace it only if you want one vendor for both jobs, for example LangSmith, Braintrust, or PostHog.

What is the best open-source alternative to Langfuse for user analytics?

Arize Phoenix ships a User Friction evaluator and a Completeness evaluator, and Opik ships UserFrustrationMetric and SessionCompletenessQuality judges for threads. PostHog's MIT core adds message sentiment, trace clustering, and funnels. Note that Phoenix uses the Elastic License 2.0, which is source-available rather than OSI-approved.

Can I send the same traces to Langfuse and another tool?

Yes. Langfuse's SDKs are built on the OpenTelemetry client and it accepts OTLP, and most of the tools on this page accept OpenTelemetry too. Agnost AI can be added as an extra span processor next to your existing exporter, so both tools receive the same spans.