Alternatives · Langfuse
6 best Langfuse alternatives in 2026
Langfuse is a strong open-source tool for tracing, evals, and prompts. Its automatic failure grouping, Topics, is "In progress" on its public roadmap, and its GitHub release notes show it in a closed beta on Langfuse Cloud in October 2026. Today, you write the judges and run the error analysis yourself. Several alternatives below find failures for you, but each needs something first, such as a paid plan, a feature turned on in each project, or a scheduled scan. three.dev needs no setup. Here is what each one is best for, what it costs, and where each one beats Langfuse.
Bryan Lee, Product LeadUpdated · 16 min read
Quick comparison: the best Langfuse alternatives
| Tool | Best for | Starting price | Key difference from Langfuse | Review |
|---|---|---|---|---|
| 01three.dev | Teams with a live AI feature and a coding agent, and no time to write evals | Free; paid from $79 a month | Does one job: finds, names, and ranks your failures by severity from live traffic. Nothing to configure. | Read the three.dev review |
| 02LangSmith | Teams on LangChain or LangGraph that want fixes drafted as pull requests and accept per-seat pricing | Free for 1 seat; Plus $39 per seat a month | Closed source and priced per seat. Engine drafts prompt and code fixes on paid plans. | Read the LangSmith review |
| 03Braintrust | Teams that want evals in CI to decide what ships, on a hosted platform | Free; Pro $249 a month | A hosted eval platform with CI checks on every pull request. Not published as open source. | Read the Braintrust review |
| 04Raindrop | Teams that want Sentry-style monitoring and Slack alerts for agents more than a prompt and eval workbench | Free; Pro $299 a month plus usage | Monitoring for agents in production, Sentry style. Issue Detection is on Pro. | Read the Raindrop review |
| 05Arize Phoenix | Teams that self-host Langfuse today and want an OpenTelemetry-native tool they run themselves | Phoenix free to self-host; Arize AX Pro $50 a month | OpenTelemetry-native, under the Elastic License 2.0. Signal, its failure grouping, is in hosted Arize AX. | Read the Arize Phoenix review |
| 06Opik | Langfuse users who want Apache 2.0 open source with failure scans added | Free; Pro $19 a month | Apache 2.0. Diagnostics groups repeated failures on request, or daily once you turn that on. | Read the Opik review |
| ··Langfuse | Teams that want open-source tracing, evals, and prompts in one tool | Free; Core $29 a month | The tool you are comparing against. |
1. three.dev
three.dev records your AI feature's live LLM traffic, scores a sample of it in the background, and groups the failures into named failure modes. In Langfuse, that work is yours: an LLM connection, a judge prompt, a rule, and error analysis by hand. three.dev needs none of it. It sets each failure mode's severity and ranks the list, most severe first, and your team can reshape the list where it disagrees. Your coding agent then reads a failure mode and its failing requests and proposes the fix.
Failure modes
Failures group into failure modes, ranked by importance and frequency.
A failure mode table counts up, each row gets a severity, and the rows sort by severity. Wrong refund window is High and leads with 386 of 8,124 scored requests. Invented discount code is Medium and climbs above Unrequested bullet lists, which is Low, though it has fewer failures.
Manage failure modes
Reshape the failure modes until the list matches your team's mental model.
A person ticks Ignored language request and Replied in English, then clicks Merge. The two rows become one, Ignored language request, with 104 of 8,124.
Where it is strong
- Nothing to write or tune. Scoring starts on your live traffic, and issues group into named failure modes once enough of them collect.
- Your team can change a severity, create a failure mode in plain language, split or merge failure modes, or dismiss one.
- Every issue says why the request failed and quotes the lines behind it, so you can check each one yourself.
- Claude Code, Codex, and Cursor read a failure mode and its failing requests over the three.dev MCP server. A Fix with your agent button starts one on a single failure mode.
Where it falls short
- Not open source, and no self-hosted version. Your LLM requests pass through the proxy at gate.three.dev.
- No prompt management, datasets, experiments, or CI gates. If you use those in Langfuse, keep it alongside.
- It records each LLM call with the tool calls the model asks for. Tool runs and retrieval steps get no spans of their own.
- No alerts. Not for health data (PHI) or government workloads yet.
- Pricing
- Free is $0 with 1 GB of data (about 50,000 requests) and 5k scored requests a month. Starter is $79 a month with 5 GB and 10k scored requests, then $3 per GB and $2.50 per 1k scored. Pro is $249 a month with 20 GB and 25k scored requests, then $2 per GB and $1.50 per 1k scored. Every plan includes scoring and unlimited seats.
- Choose it over Langfuse if
- you have a live AI feature, nobody has time to write and tune judge prompts in Langfuse, and you want your coding agent to fix the most severe failures first.
Why teams choose three.dev
"Using three.dev has been a huge leap for us in providing our users the highest quality while maintaining fast responses."
Don PottingerCo-founder and CTO, LanguaTalkRead the case studythree.dev vs Langfuse
| Feature | three.dev | Langfuse |
|---|---|---|
| What it is | Failure-mode discovery for live AI features, with the fix handed to your coding agent | An open-source platform for tracing, evals, prompt management, datasets, and experiments |
| How you connect | Change one base URL, add a three.dev API key and one header. Your AI SDK stays the same. | Add the Langfuse SDK (Python or TypeScript) or OpenTelemetry. 100+ integrations. |
| Setup before you see failures | None. Scoring starts on your live traffic. | Add an LLM connection, write a judge prompt (or start from a template), and attach a rule that picks which observations to score. |
| Grouping failures | Automatic. Issues group into named failure modes, each with a severity, most severe first. | You run error analysis, or ask the Halo assistant to cluster errors. Automatic grouping (Topics) is "In progress" on the public roadmap and in closed beta on Langfuse Cloud (October 2026). |
| Shaping the list | three.dev sets each severity, and your team can change it. Your team can also create a failure mode in plain language, split or merge failure modes, or dismiss one. | During error analysis, you cluster failures into named categories and refine the names yourself. An LLM can draft the first list. |
| Asking about your traffic | Chat: ask in plain words, and each answer links to the requests behind it. | Halo: explores your project data and takes actions in Langfuse with your approval. |
| Coding-agent access | The MCP server hands over a named failure mode, its failing requests, and why each one failed. A Fix with your agent button copies the prompt. | An MCP server with read and write tools, a CLI, an agent skill, and a Cursor plugin give your agent traces, scores, prompts, and datasets. |
| Checking a fix | Tag requests with a release, then compare two releases or two date ranges on quality, latency, and cost. | Releases and versions track cost, latency, and quality, and experiments compare runs on a dataset. |
| What gets recorded | Each LLM call, including the tool calls the model asks for, grouped into sessions | Nested traces: LLM calls, tool calls, and retrieval steps |
| Prompts, datasets, experiments | Not offered today | Yes, on every plan |
| Alerts | Not offered today | Alerts on observation metrics and scores, to Slack, webhooks, or GitHub Actions |
| Open source and self-hosting | Hosted proxy at gate.three.dev | MIT-licensed core, free to self-host |
| Price | Free plan; Starter $79, Pro $249 a month. Scoring included. Unlimited users. | Free Hobby plan; Core $29, Pro $199, Enterprise $2,499 a month; self-hosting free |
What you gain moving from Langfuse
- Failure modes found for you, with no judge prompts to write or tune
- A list of what to fix first, ranked by severity, from your real traffic
- A list your team edits in place: change a severity, create a failure mode in plain words, split, merge, or dismiss
- A coding agent that starts from a named failure mode and its failing requests
- Setup that keeps your AI SDK: one base URL, one key, one header
What you give up
- Open source and self-hosting
- Prompt management, datasets, experiments, and CI gates
- Alerts on metrics and scores
- Nested traces of tool calls and retrieval steps
2. LangSmith
LangSmith is LangChain's agent engineering platform: tracing, online and offline evals, prompts, alerts, and deployment. Leaving Langfuse for it means leaving open source and free self-hosting behind. LangSmith is closed source, and self-hosting is an add-on to its Enterprise plan. In return, you get Engine. On Plus and Enterprise, it scans your traces on a schedule, groups failures into issues ranked by severity, and drafts prompt and code fixes as GitHub pull requests or GitLab merge requests.
Where it is strong
- Engine drafts prompt and code fixes, opens them for review on GitHub or GitLab, and pushes a revision when its own CI checks fail.
- Made by the LangChain team, with native tracing for LangChain and LangGraph and SDKs for Python, TypeScript, Go, and Java.
- Multi-turn online evaluators score a whole conversation, where Langfuse evaluators read one observation at a time.
- Deployment lives in the same platform, so one vendor covers build, test, deploy, and monitor.
Where it falls short
- Closed source. Self-hosting is an Enterprise add-on.
- Plus costs $39 per seat a month, invited users count as seats, and removed seats earn no credit.
- Engine bills in usage units on top of seats and traces: about 7 to 45 units of $1 per run, with a default cap of 750 units a month. Insights and Engine are not on the free plan.
- Base traces keep 14 days, and online evaluators upgrade matching traces to paid 180-day retention by default.
- Pricing
- Developer is free for 1 seat and 5k base traces a month. Plus is $39 per seat a month with 10k base traces, then $0.005 per trace, with Engine usage billed on top at $1 per unit. Enterprise is custom and adds self-hosting. LangSmith bills traces and seats, where Langfuse bills every trace, observation, and score as a unit.
- Choose it over Langfuse if
- your app runs on LangChain or LangGraph and you want fix pull requests drafted for you, enough to give up open source and pay per seat.
3. Braintrust
Braintrust calls itself the active observability platform for agents. For a Langfuse user, much of it is familiar: tracing, datasets, scorers, experiments, prompts, and alerts. Topics clusters production traces by task, sentiment, and issues once an owner turns it on in each project. Patterns, in public preview, looks for recurring problems on a schedule. Braintrust bills processed data, scores, and model credits.
Where it is strong
- A GitHub Action runs evals on each pull request, posts the results, and can gate merges.
- Loop drafts scorers, prompts, and datasets, waits for your approval before it changes anything, and can open GitHub pull requests (public preview).
- Topics sorts every trace by what the user asked for and how they felt, as well as by issues, with built-in facets.
- SDKs in six languages (TypeScript, Python, Go, Java, Ruby, and C#), OpenTelemetry, and alerts to Slack, webhooks, or PagerDuty.
Where it falls short
- The platform is not published as open source. Self-hosting and BYOC are Enterprise only, and Braintrust still runs the control plane.
- No plan sits between free Starter and Pro at $249 a month, and Starter keeps logs for 14 days.
- A member of the organization's Owners group must turn Topics on in each project, and clusters appear only after 100 facet summaries.
- Topics and Loop run on monthly model credits that do not roll over. On Starter without a card on file, Topics pauses when the credit runs out.
- Pricing
- Starter is free with 1 GB of processed data, 10k scores, and $10 of model credits a month. Pro is $249 a month with 5 GB, 50k scores, and $100 of credits. Enterprise is custom. Processed data is measured at ingestion, so deleting logs does not lower it.
- Choose it over Langfuse if
- you want evals to decide what ships, with CI checks on every pull request, and you no longer need to self-host.
4. Raindrop
Raindrop is a monitoring platform for AI agents that works the way Sentry does for web apps: tracing, search, signals, alerts, and experiments. Langfuse is a workbench for traces, prompts, and evals. Raindrop watches production and tells you when something changes. Signals track behaviors you describe in plain language or pick from presets such as user frustration and task failure. On Pro, Issue Detection looks for failures spreading across users, new breakage patterns, and regressions after a deploy.
Where it is strong
- Issue Detection flags regressions after a deploy and failures spreading across users, and shows each issue's trend.
- Real-time Slack alerts when a signal crosses a threshold.
- Seven preset signals, from user frustration and refusals to jailbreaks and praise, and presets use none of your custom slots.
- Experiments compare cohorts of events and measure how signal rates change.
Where it falls short
- Issue Detection, Stumbles (one-off failures), and Experiments need Pro, at $299 a month plus $0.003 per event.
- The free Hobby plan stops recording for the rest of the month after 1,000 events.
- Custom signals are capped: 1 on Hobby and 10 on Pro.
- Hosted. Self-hosting in your own cloud is in beta.
- Pricing
- Hobby is free with 1,000 events a month and 14 days of retention. Pro is $299 a month plus $0.003 per event up to 1M, then $0.002, with unlimited retention and a 14-day free trial. Enterprise is custom. One event is a user message and the agent's response, including tool calls and sub-agents, or one background agent run.
- Choose it over Langfuse if
- you want alerts when agents break in production more than a place to manage prompts and evals, and a $299 monthly base fits.
5. Arize Phoenix
Like Langfuse, Phoenix is an open-source tool for tracing, evals, datasets, experiments, and prompts that you can self-host for free. It is built on OpenTelemetry and uses the Elastic License 2.0, where Langfuse's core uses MIT. Arize AX, the hosted platform, adds online evals on live traffic, alerts, and Signal, which groups recurring production failures into ranked issues with a proposed fix. Dynatrace completed its purchase of Arize on October 1, 2026.
Where it is strong
- Free to self-host with no license fees, usage limits, or feature gates, and it can run fully air-gapped.
- Built on OpenTelemetry, with instrumentation for many model providers and agent frameworks.
- Eval templates or your own judges, code evaluators, and CI gates for pytest, Vitest, or Jest.
- In AX, a Signal issue can become a running evaluator: Alyx builds one for that failure pattern and starts scoring new traces.
Where it falls short
- The Elastic License 2.0 bars offering Phoenix to others as a hosted service.
- Phoenix's docs send you to Arize AX for online evals on live traffic and for alerting, and Signal is part of AX.
- AX Free and Pro cap Signal at 10 and 25 issues a month, and fix pull requests are Enterprise only.
- Ownership just changed. Dynatrace says it will keep supporting Phoenix and AX, and fold Arize into its own platform over time.
- Pricing
- Phoenix is free to self-host; you run the servers. Arize AX Free gives 25k spans and 1 GB a month, with 15 days of retention. AX Pro is $50 a month with 50k spans and 10 GB, then $0.0008 per span. Enterprise is custom and adds self-hosted AX.
- Choose it over Langfuse if
- you self-host Langfuse today, want to stay on open source you run yourself, and want OpenTelemetry-native tracing.
6. Opik
Opik is Comet's open-source platform for tracing, evals, prompt management, and production monitoring, under the Apache 2.0 license. It sits close to Langfuse: open source, self-hosted on a similar stack, with a lower cloud price. It adds Diagnostics, run by the Ollie assistant. Diagnostics scans your traces on request, or every day once you switch that on, groups repeated failures into one issue each, and suggests a root cause and a fix.
Where it is strong
- Apache 2.0, with one codebase for the open-source and hosted versions.
- Online evaluation, alerts, annotation queues, and a prompt library come in the free open-source build.
- The lowest paid plan on this list: Pro is $19 a month for 100k spans, then $5 per 100k.
- AI guardrails come with the self-hosted build.
Where it falls short
- Diagnostics and Ollie run on credits, and Comet does not publish the credit price.
- Daily scans are off until someone turns them on in each project. They cover only the previous day, and missed days are not scanned later.
- SSO and RBAC are Enterprise only, and self-hosted Opik has no user management. Langfuse's free self-hosted build includes SSO and RBAC.
- Self-hosting means running ClickHouse, MySQL, Redis, and object storage.
- Pricing
- Self-hosted Opik is free. Free Cloud gives 25k spans a month with 60 days of retention. Pro is $19 a month for 100k spans, then $5 per 100k spans. Enterprise is custom. Ollie credits are extra, at a price Comet does not publish.
- Choose it over Langfuse if
- you want an Apache 2.0 tool you can self-host, with guardrails and online evals in the free build, or a $19 cloud plan.
Which tool is right for you?
- You have no eval setup and want failures found, ranked, and handed to your coding agentthree.dev
- Your app runs on LangChain or LangGraph and you want fixes drafted as pull requestsLangSmith
- You want evals and CI checks on every pull requestBraintrust
- You want Sentry-style alerts on agents in productionRaindrop
- You self-host today and want OpenTelemetry-native tracingArize Phoenix
- You want an Apache 2.0 tool you can self-host, or a $19 cloud planOpik
Questions
What is the best Langfuse alternative?
It depends on why you are leaving. three.dev fits teams with no eval setup that want failures found and ranked for them. LangSmith fits LangChain and LangGraph apps. Braintrust fits teams that run evals in CI. Raindrop fits teams that want alerts on agents in production. Arize Phoenix and Opik fit teams that must self-host open source.
Which Langfuse alternatives find production failures automatically?
three.dev needs no setup. Send requests through it, and it scores a sample of them in the background and reports each defect it finds as an issue. Once enough issues collect, it groups them into named failure modes, ranked by severity. The other tools each name a condition in their own docs, as of October 29, 2026. LangSmith Engine needs a Plus or Enterprise plan, an admin to turn it on, and a switch in each tracing project. It scans every 6 hours, per LangSmith's pricing page, and each run uses about 7 to 45 usage units at $1 each, up to a default cap of 750 units a month. Braintrust Topics needs a member of the organization's Owners group to turn it on in each project. It uses built-in facets (task, sentiment, and issues) plus any custom facets you write, shows clusters after 100 facet summaries, runs daily, and uses model credits that do not roll over. Braintrust Patterns, in public preview, runs daily once someone turns it on in a project. Raindrop Issue Detection needs the Pro plan, at $299 a month plus $0.003 per event. Arize AX Signal is turned on for each project, scans every 6 hours by default, stops at 10 issues a month on Free and 25 on Pro, and is not in Phoenix. Opik Diagnostics needs the Ollie assistant and credits whose price Opik does not publish, and it scans daily only after someone turns that on. Langfuse itself lists Topics as "In progress" on its public roadmap, and its GitHub release notes show Topics in a closed beta on Langfuse Cloud, for allowlisted projects, in October 2026.
Is there a free Langfuse alternative?
Arize Phoenix and Opik are free to self-host. three.dev, LangSmith (1 seat), Braintrust, Raindrop, Arize AX, and Opik Cloud have free cloud plans with usage limits. three.dev's Free plan covers about 50,000 requests and 5,000 scored requests a month, with unlimited seats. Langfuse counts each trace, observation, and score as a unit, so the same traffic is more than the 50,000 units a month in its free Hobby plan. On Langfuse Cloud it needs at least Core at $29 a month, plus your model provider's bill for the calls that score requests.
Which Langfuse alternatives are open source?
Opik is Apache 2.0, and Arize Phoenix is under the Elastic License 2.0. Both are free to self-host. LangSmith is closed source, and Braintrust does not publish its platform as open source. Both self-host only on Enterprise plans, and Braintrust still runs the control plane. Raindrop is hosted, with self-hosting in beta. three.dev is a hosted service and is not open source.
Langfuse vs LangSmith: which is better?
Langfuse is open source and free to self-host. LangSmith is closed source and self-hosts only on Enterprise. Langfuse's paid cloud plans and every three.dev plan have unlimited users; LangSmith Plus charges $39 per seat. LangSmith ships Engine on Plus and Enterprise, billed per run: it groups failures into issues and drafts fix pull requests. Langfuse's automatic grouping (Topics) is in closed beta. Pick Langfuse for open source and free self-hosting. Pick LangSmith for LangChain depth and drafted fixes.
Arize vs Langfuse: which is better?
Both have a build you can self-host for free. Langfuse's core is MIT licensed. Arize Phoenix uses the Elastic License 2.0, which bars offering it as a hosted service, and it is built on OpenTelemetry. Arize's hosted platform, AX, adds Signal, which groups recurring failures into ranked issues, with caps of 10 issues a month on Free and 25 on Pro. Langfuse's automatic grouping (Topics) is in closed beta. Pick Langfuse for one MIT tool for tracing, prompts, and evals. Pick Arize for OpenTelemetry-native tracing and evals that gate CI in pytest, Vitest, or Jest.
Is Langfuse still open source after ClickHouse bought it?
Yes. ClickHouse acquired Langfuse in January 2026. Langfuse says there are no planned licensing changes, its core is still MIT licensed, and you can still self-host it for free.
Can I use three.dev and Langfuse together?
Yes. three.dev only changes the base URL your AI SDK calls, so your Langfuse tracing can stay in place. You can keep Langfuse for traces and prompts and add three.dev to find and rank your failures.
How do I move from Langfuse to three.dev?
Point your AI SDK at gate.three.dev, add a three.dev API key, and add the X-Three-Use-Case header. Or paste one setup prompt into your coding agent and let it make the change. Scoring then starts on your live traffic.
Does three.dev slow down my requests?
No. Scoring runs after the response returns, so it never changes a request or adds latency.
How we made this list
We make three.dev, so we list it first. For every other tool, we read its own site, docs, pricing page, and changelog on October 29, 2026, and each review names where that tool is stronger than three.dev. Prices and features change often, so check each pricing page before you buy.
Sources
- three.dev docs: sending requests
- three.dev docs: live scoring
- three.dev docs: failure modes
- Langfuse pricing
- Langfuse self-hosted pricing
- Langfuse roadmap
- Langfuse: Topics in closed beta (GitHub)
- Langfuse: online evaluation
- Langfuse: LLM-as-a-judge
- Langfuse billable units
- Langfuse Halo
- Langfuse alerts
- Langfuse joins ClickHouse
- LangSmith pricing
- LangSmith Engine
- LangSmith Engine setup
- LangSmith Insights
- LangSmith usage and billing
- LangSmith self-hosting
- LangSmith changelog
- Braintrust pricing
- Braintrust plans and limits
- Braintrust Topics
- Braintrust: enable Topics
- Braintrust Patterns
- Braintrust deployment options
- Braintrust self-hosting
- Raindrop pricing
- Raindrop plans and limits
- Raindrop Issue Detection
- Raindrop signals
- Arize pricing
- Arize AX Signal
- Phoenix self-hosting
- Phoenix license
- Dynatrace completes Arize acquisition
- Opik pricing
- Opik Diagnostics
- Opik self-hosting
Ship on evidence, not vibes.
See how your agent fails. Hand the fix to your coding agent.