Langfuse vs Helicone: Tracing SDK or Proxy Gateway?
November 19, 2026
Before you compare features, check Helicone's status. As of November 19, 2026, Helicone's own homepage carries the banner "Helicone is now part of Experiential Labs. Details to come soon." Earlier this year Mintlify acquired Helicone and said it would run in maintenance mode. Neither company has published what the new ownership means for the roadmap.
Langfuse is in a steadier spot. It joined ClickHouse in January 2026, said there are no planned licensing changes, and is still shipping, including a new AI gateway in public beta.
So the Langfuse vs Helicone decision looks like this:
- Pick Langfuse if you want tracing, evals, prompt management and datasets in one open-source platform you can self-host, and a vendor with a clear roadmap.
- Pick Helicone if you already run it, or you specifically want a proxy-style gateway with edge caching and per-request cost tracking set up by changing one URL, and you can live with an open ownership question.
- The core design differs. Helicone is built around a proxy gateway that logs every model call. Langfuse is built around SDK and OpenTelemetry tracing of your whole app, LLM calls, tool calls and retrieval steps included.
This guide uses only what each vendor's own site, docs, GitHub repo and pricing page say today. It skips vendor-written "X vs Y" pages and third-party benchmarks.
Langfuse vs Helicone: Quick Comparison
| Langfuse | Helicone | |
|---|---|---|
| Status, Nov 2026 | Part of ClickHouse since January 2026, actively shipping | Banner says "now part of Experiential Labs"; earlier put in maintenance mode by Mintlify |
| Licence | MIT, except the ee folders | Apache-2.0 |
| Integration style | SDKs and OpenTelemetry; AI gateway in public beta | Change the base URL to its AI gateway |
| Tracing depth | Nested traces of LLM calls, tools, retrieval and agent steps | Requests, sessions and user-level views |
| Evals | LLM-as-a-judge, code evaluators, human labels, datasets, experiments | Datasets, playground and prompt tools |
| Prompt management | Versioned prompts with deploy and rollback | Versioned prompts deployed through the gateway |
| Response caching | Not described in the docs we read | Header-based caching at the edge |
| Self-hosting | Docker Compose, Kubernetes (Helm), Terraform for AWS, Azure, GCP | All-in-one Docker image; Helm for enterprise |
| What gets billed (cloud) | Billable units: traces, observations, scores | Requests, plus storage |
Helicone's Status in November 2026
Here is the timeline from primary sources.
March 3, 2026. Mintlify announced it was acquiring Helicone. The post said Helicone "will continue operating in maintenance mode" with security updates, bug fixes and new models still shipping, and that Mintlify would help customers move "to another platform."
Now. Helicone's homepage says it is now part of Experiential Labs, with details to come. Experiential Labs is a Y Combinator company building an open-source AI gateway. Neither company has published the terms, so we won't guess at them.
What we could confirm on November 19, 2026:
- The Helicone GitHub repo is public, Apache-2.0 and not archived. Its latest commit is dated November 17, 2026.
- The Helicone pricing page still lists plans with sign-up buttons.
- The docs are online, with no deprecation notice on the caching and self-hosting pages.
What this means for you. If Helicone already runs in production, nothing has been switched off, and the code stays Apache-2.0 if you self-host. If you're choosing a tool for the next two years, the last public roadmap statement was "maintenance mode," and the new owner hasn't said what comes next. Factor that in.
Integration: Proxy vs SDK
This is the practical difference you'll feel on day one.
Helicone sits between your app and the model provider. Its README shows the pattern: point your OpenAI client's baseURL at Helicone's AI gateway and use a Helicone API key. The gateway gives you 100+ models behind one key, with routing and fallbacks, and logs every request on the way through. No new SDK, no code around each call.
The upside is speed: logging in minutes. The downside is depth. A proxy sees the request and the response. It doesn't see the retrieval step that came before or the tool call your agent made afterwards, unless you tag requests with session and custom properties.
Langfuse instruments your app. You add its Python or TypeScript SDK, or send OpenTelemetry spans from any language. The Langfuse GitHub README lists 100+ integrations, including LangChain and LangGraph, LlamaIndex, the Vercel AI SDK, the OpenAI Agents SDK, the Claude Agent SDK, Pydantic AI and LiteLLM. Each trace nests LLM calls, tool calls, retrieval and agent steps, so you can see where a bad answer actually went wrong.
Langfuse now has a gateway too, but it's early. The Langfuse AI Gateway went into public beta on November 4, 2026. You change the base URL and API key, and Langfuse forwards requests and records each call. During the beta it supports OpenAI and Anthropic only, and self-hosted users can run the open-source gateway next to Langfuse. The announcement doesn't mention caching or fallbacks, so don't treat it as a like-for-like Helicone replacement yet.
| You want | Better fit |
|---|---|
| Logging with zero code changes, many providers | Helicone gateway |
| Full traces of agents, tools and retrieval | Langfuse SDK or OpenTelemetry |
| One base-URL change, OpenAI or Anthropic only | Either (Langfuse gateway is in beta) |
| Vendor-neutral instrumentation | Langfuse via OpenTelemetry |
Tracing, Evals and Prompt Management
Langfuse goes deeper on quality. Its platform covers:
- Tracing with filters by user, session, cost, latency or metadata.
- Evaluations using LLM-as-a-judge, code evaluators, user feedback and manual labelling, on production traces or in experiments.
- Datasets and experiments to run test cases and compare versions side by side.
- Prompt management with versions, labels, one-click deploys and rollbacks, plus a playground that replays real inputs.
- Dashboards, a metrics API and alerts.
If you're asking "did this prompt change make answers better?", Langfuse has more built in.
Helicone is strongest on operations. Its dashboard centres on requests, sessions, users and segments, with cost and latency per request, rate limits and alerts, and a query language (HQL) for slicing logs. It also has prompts, datasets and a playground, and exports analytics to PostHog. That covers "what is this costing us and which user is causing it?" very well, with less on structured evaluation.
Caching and Cost Control
Response caching is a Helicone strength. The Helicone caching docs describe it: send a Helicone-Cache-Enabled: true header and identical requests return the stored response. Responses live in Cloudflare Workers KV across edge locations, with a default lifetime of seven days set through Cache-Control. You can store several responses per request, namespace the cache per user with a seed, and ignore selected JSON keys when matching. For repeated prompts in testing or high-traffic apps, that cuts both latency and spend.
Langfuse tracks cost, but response caching isn't its job. It records token usage and cost per generation, and you can filter traces by user, session, cost or latency, which is enough to find expensive prompts. Its new gateway adds central provider keys and spend tracking, but the beta announcement doesn't describe response caching. If you need a cache, you'd put one in your own code or in a separate gateway.
Self-Hosting and Licence
Both are open source. The work to run them differs.
Langfuse self-hosts on Docker Compose for a single machine, or on Kubernetes with Helm, which it calls the preferred production deployment, plus Terraform templates for AWS, Azure and GCP. The self-hosting docs list the parts you run: Langfuse Web and Worker containers, Postgres, ClickHouse, Redis or Valkey, and S3-compatible blob storage. The repo is MIT licensed except the ee folders. A few add-ons, such as organization creators, the instance management API and UI customization, need a licence key when self-hosted. Self-hosted instances send telemetry by default, and you can switch it off.
Helicone self-hosts with a single all-in-one Docker image that bundles the dashboard, its Jawn API, Postgres, ClickHouse and MinIO storage. Two notes from its Docker docs: the Jawn service no longer proxies LLM traffic, so you deploy the AI gateway alongside it, and the proxy port has no authentication, so you must firewall it. Helm charts are available to enterprise customers. The code is Apache-2.0.
| Langfuse | Helicone | |
|---|---|---|
| Quick start | docker compose up | One docker run command |
| Production path | Helm on Kubernetes, Terraform | Helm (enterprise), gateway deployed separately |
| Data stores | Postgres, ClickHouse, Redis/Valkey, S3 | Postgres, ClickHouse, MinIO |
| Licence | MIT, except ee folders | Apache-2.0 |
| Ownership risk | Low; Langfuse says no planned licensing changes | Open; new owner, no details yet |
The Langfuse announcement on joining ClickHouse says Langfuse stays open source and self-hostable, and that Langfuse Cloud keeps running as before.
Cloud Pricing Models
Both clouds have a no-cost entry tier and usage-based paid plans. They count different things.
Langfuse Cloud bills units: every trace, observation (spans, events, generations) and score you send counts. The Langfuse pricing page lists a Hobby tier, paid Core and Pro plans that include a block of units with graduated overage, and Enterprise. Longer data retention and team features sit on the higher plans. Self-hosting the open-source version costs you only your infrastructure.
Helicone bills requests and storage. Its pricing page, linked above, lists a Hobby tier, Pro and Team plans with usage-based pricing on top, and Enterprise.
| Your workload | Watch out for |
|---|---|
| Deep agent traces, many spans per request | Langfuse units grow with every span |
| Many small, simple model calls | Helicone request counts |
| Large prompts and responses | Helicone storage |
| Strict data residency | Self-host either one |
Prices change. Check both pages before you budget.
Langfuse vs Helicone: Which Should You Pick
Choose Langfuse if:
- You're building agents or RAG pipelines and need to see every step, not just the model call.
- Evaluations, datasets and prompt versioning matter as much as logs.
- You want a vendor with a clear, active roadmap and an MIT core.
- You want to instrument once with OpenTelemetry and keep your options open.
Choose Helicone if:
- It already works in production and you're happy to watch the ownership news.
- You want logging and cost tracking from a one-line base-URL change across many providers.
- Edge response caching is a real cost saver for your traffic.
- You'll self-host the Apache-2.0 code and accept owning its upkeep.
If you're starting fresh today, Langfuse is the safer default. Helicone's feature set is fine; its roadmap is the unknown.
If neither fits, we cover other tools in Langfuse alternatives and Helicone alternatives. If LangChain's platform is on your list, see Langfuse vs LangSmith, and for proxy-first options, see Portkey alternatives.
Where Taku Fits
Tracing and evals are tools for the people building AI products. If you'd rather run an AI tool someone has already built than instrument your own, that's where Taku comes in. Taku Apps and Sites, built by the best AI minds. Find one, run it, and keep it in your Mac Dock. Taku is in Beta, and the Mac app is available now.
Key Points
- Helicone's ownership is in flux. Its homepage says it is now part of Experiential Labs, after Mintlify put it in maintenance mode in March 2026. The code is still Apache-2.0 and was updated on November 17, 2026.
- Langfuse joined ClickHouse in January 2026 with no planned licensing changes, and launched an AI gateway in public beta on November 4, 2026.
- Helicone is a proxy first: one base-URL change, many providers, edge caching and per-request cost.
- Langfuse is a tracing platform first: nested traces, evals, datasets and prompt management through SDKs and OpenTelemetry.
- Both self-host. Langfuse runs more parts in production; Helicone has a one-container quick start.
FAQ
What is the difference between Langfuse and Helicone?
Helicone logs model calls by sitting in front of the provider as a proxy gateway. Langfuse traces your whole application through SDKs and OpenTelemetry, and adds evals, datasets and prompt management. Helicone is quicker to set up. Langfuse shows more of what happened inside an agent.
Is Helicone shutting down?
Neither Helicone nor Mintlify has announced a shutdown. Mintlify put it in maintenance mode in March 2026. As of November 19, 2026, Helicone's homepage says it is now part of Experiential Labs, with details to come. The repo is active and the pricing page still takes sign-ups.
Is Langfuse open source?
Yes. The Langfuse repository is MIT licensed, except the ee folders. A few self-hosted add-ons need a licence key. Langfuse joined ClickHouse in January 2026 and said there are no planned licensing changes.
Does Langfuse have a gateway like Helicone?
Yes, but it's newer. The Langfuse AI Gateway entered public beta on November 4, 2026 and supports OpenAI and Anthropic during the beta. Helicone's gateway covers many more providers and includes response caching.
Can I self-host Langfuse or Helicone?
Both. Langfuse runs on Docker Compose for testing and Kubernetes with Helm for production, with Postgres, ClickHouse, Redis and S3 storage. Helicone has an all-in-one Docker image, with the AI gateway deployed alongside it.
Can I migrate from Helicone to Langfuse?
Yes, but it's a change of integration, not a config switch. You replace the base-URL proxy with Langfuse's SDK, an OpenTelemetry exporter or, for OpenAI and Anthropic calls, the Langfuse gateway in beta. Plan for a cache of your own if you relied on Helicone's caching.