Langfuse vs Helicone: tracing platform or AI gateway?

Picture two developers on the same Monday morning. One adds an SDK to their agent code so every step, prompt and score lands in a trace they can study later. The other changes a single base URL so every request to a model passes through a gateway that logs it on the way. Both end up with a record of what their AI app did. They got there in completely different ways, and that difference is the whole story of this comparison.

For readability, I'll call the first tool the tracing platform and the second one the gateway tool. That's how each one describes itself at heart, even though both have started borrowing from the other's side.

Two ways to get your model calls on the record

The tracing platform puts instrumentation first. You add its SDKs or send data over OpenTelemetry, and tracing, evals, prompts and datasets sit at the centre of the product. Its site describes it as one open platform for every AI call. It's built for teams who want to dig into why an agent did something, not just that it did it.

The gateway tool flips that around. You send requests through an OpenAI-compatible endpoint and they're logged automatically, with provider routing, caching, custom rate limits and fallbacks on top. The vendor says it's tuned for fast set-up, and the logic makes sense: if the gateway is already in the path of every call, there's nothing extra to instrument.

So the first question isn't which dashboard looks nicer. It's whether you want to change your code or change your endpoint.

Side by side, one topic at a time

What differs	Langfuse	Helicone
Core approach	SDKs or OpenTelemetry; tracing, evals, prompts and datasets at the centre	Gateway first: requests through an OpenAI-compatible endpoint are logged automatically
Open source and hosting	Open source, with Docker Compose, Kubernetes and Terraform self-hosting guides for AWS, GCP and Azure	Docs call it OSS LLM observability and include a self-hosting section
AI gateway	Newly launched AI Gateway, labelled beta in its docs, with gateway keys and model access	The headline product, with provider routing, caching, custom rate limits and fallbacks
Evals and datasets	Evals, datasets and experiments alongside tracing	Datasets, eval scores and user feedback listed under observability
Prompt management	Prompt and Skill management is a top-level product	A docs section covering prompt assembly and SDK integration
Agent and tooling extras	In-app agent called Halo, plus MCP, CLI and API access to every feature	HQL query language and a PostHog integration in its docs
Company status	Says it has joined ClickHouse	Banner says it's no longer part of Mintlify, with details to come

Who owns these tools now, and why you should ask

This is the part most comparisons skip, and it's the one I'd weigh most for a long-term choice. Both companies have had a change of ownership recently, and both say so on their own home pages.

The tracing platform says it has joined ClickHouse. The gateway tool's home page carries a banner saying it's no longer part of Mintlify, with more details promised. Neither page sets out what happens next for the product roadmap, so I won't guess. What I'd say is simple: if you're about to put one of these in the path of every production call, ask the vendor directly what the change means for support and for the open source project.

The open source side does soften the risk. Both describe themselves as open source and both document self-hosting, so you're not locked into a hosted service if things shift. That's a real comfort, but it's only worth something if your team is willing to run the thing itself.

Self-hosting, prompts and evals in practice

Running it on your own servers

The tracing platform goes further on this front. Its self-hosting guides cover Docker Compose, Kubernetes and Terraform, with routes for AWS, GCP and Azure. The gateway tool's docs have a self-hosting section too, but its pages don't spell out the same range of options, so I'd read the section before assuming it matches.

Prompts as a product, or as a feature

If your team edits prompts often, notice how each one ranks them. The tracing platform treats Prompt and Skill management as a top-level product. The gateway tool covers prompt management in a docs section with prompt assembly and SDK integration. Both can do the job; one simply puts it closer to the front door.

Evaluation before you ship

Evals, datasets and experiments sit right next to tracing on the tracing platform, which suits teams that score agent output as part of their release routine. The gateway tool lists datasets, eval scores and user feedback under observability. Again, it's there, but it isn't the headline.

There's one more twist worth knowing. The tracing platform has launched its own AI Gateway, labelled beta in its docs. So it's moving towards the gateway tool's home ground, while the gateway tool already offers evals and prompts. The overlap is growing, which makes the core approach matter even more.

A quick checklist before you choose

• Will your engineers add SDK or OpenTelemetry instrumentation, or do they want a one-line endpoint change?
• Do you need caching, fallbacks and rate limits on model calls today? That points to the gateway tool.
• Is prompt and eval work a daily habit for your team? That points to the tracing platform.
• Must the whole thing run on your own cloud? Read both self-hosting guides side by side.
• Have you asked each vendor what its recent ownership change means for you?

If you're still mapping the wider field, our round-up of the LLM observability tools worth a look covers more options, and there's a separate page of other gateways and loggers like Helicone if routing is what you're after.

My advice: decide how you'll wire it in, then compare features

Pick the tracing platform if you want open source tracing, evals and prompt management you can self-host and feed through OpenTelemetry, and you don't mind instrumenting your code. Pick the gateway tool if the fastest route is a gateway in front of your model calls, with logging, caching and fallbacks from day one. Neither vendor crowns itself over the other, and I won't either. Choose on instrumentation versus gateway, then check each company's ownership news before you commit.

Three more LLM observability tools to consider

LangSmith

Best for: Observability, evaluation and deployment for agents and LLM apps, and a natural pick for teams already building with LangChain tooling.

Arize AI

Best for: Agent observability, evaluation and improvement in one platform; the vendor says it finds issues and proposes fixes.

Braintrust

Best for: Teams that treat evals as part of release testing, with what the vendor calls active observability that traces everything and tests what ships.

Do you make an LLM observability or AI gateway tool? Tool makers can ask to be considered for a featured place on this page.

Want your tool on this list? Featured spots and reviews here are paid partnerships. Here's how to partner with me.

Frequently asked questions

Is Langfuse open source?

Yes, it describes itself as open source and offers self-hosting guides for Docker Compose, Kubernetes and Terraform, alongside a cloud option.

Does Helicone need code changes?

Very few. It's gateway first: you send requests through an OpenAI-compatible endpoint and they're logged automatically.

Which one has an AI gateway?

Both now do. The gateway tool's gateway is its headline product. The tracing platform's AI Gateway is newly launched and labelled beta in its docs.

What's the ownership news on each?

The tracing platform says it has joined ClickHouse. The gateway tool's banner says it's no longer part of Mintlify, with details to come.

Which is better for prompt management?

The tracing platform makes Prompt and Skill management a top-level product. The gateway tool covers it in a docs section with prompt assembly and SDK integration.