What should Helicone users move to? Mintlify acquired Helicone on 3 March 2026 and said it would keep operating in maintenance mode, with security updates, bug fixes and new models still shipping, and would help every customer migrate1. Its sign-up page now reads "Sign ups are disabled"2. Choose the replacement by the job Helicone did for you: Langfuse for tracing, evaluation and prompt management; LiteLLM or Bifrost for a self-hosted gateway with cost tracking; a BYOK control plane such as INFRO for a hosted gateway that enforces budgets and attributes cost on your own provider accounts.
Nothing is being switched off, so this is a migration to plan, not an emergency. Plans made calmly keep the history you were keeping Helicone for.
What Helicone did, and which part you depend on
Helicone grew from a logging proxy into an observability platform with a gateway: request logging and tracing, cost tracking, user and property tags, evaluations and prompt management. Today its quick start points the OpenAI SDK at https://ai-gateway.helicone.ai with a Helicone API key3. Most teams use a subset. Write yours down before you compare products:
| If you mainly use Helicone for… | Look at | Because |
|---|---|---|
| Traces, sessions, evaluations, prompt versions | Langfuse | Open-source tracing, prompt management and evaluation, self-hostable |
| A self-hosted gateway with spend tracking | LiteLLM, Bifrost | Open-source gateways with virtual keys and budgets |
| Cost per request and per customer, with limits that act | A BYOK control plane (INFRO) | Hosted gateway with budgets before the call and attribution |
| Exactly Helicone, self-hosted | Helicone's own repository | Apache 2.0, Docker documented; maintenance-mode horizon |
The alternatives
Langfuse: tracing and evaluation
Langfuse describes itself as an open-source AI engineering platform for debugging, analysing and iterating on AI applications. It records traces of LLM and non-LLM calls, sessions and agent graphs; versions prompts with label-based deployment; runs LLM-as-a-judge and human evaluation with annotation queues and experiments; and tracks cost and usage per user. It is open, self-hostable and extensible4. Langfuse records what your application sends it rather than sitting in the request path; its roadmap lists a "thin, bring-your-own-key gateway with virtual keys and access controls" and calls the roadmap "directional, not a commitment"5.
Best when Helicone was your observability and evaluation tool and routing is solved elsewhere.
LiteLLM: the default self-hosted gateway
LiteLLM is a self-hosted, OpenAI-compatible gateway for more than 100 providers, with virtual keys carrying their own models, budgets and rate limits, spend tracking by key, team, user and tag, load balancing and fallbacks, guardrails, and logging to tools such as Langfuse, Datadog and Prometheus. It is deployed with Docker, Helm or Terraform against Postgres6.
Best when you used Helicone as a proxy for cost tracking and are willing to run the gateway yourself. LiteLLM plus Langfuse is a common pairing that covers both halves of what Helicone did.
Bifrost: self-hosted, with an MCP gateway
Bifrost is Apache 2.0, starts with npx or Docker, and lists automatic failover, load balancing across keys and providers, virtual keys with hierarchical budgets and rate limits, semantic caching, an MCP gateway and Prometheus metrics with tracing7.
Best when you want a self-hosted gateway with budgets and an MCP gateway in one binary.
Helicone itself, self-hosted
The repository is public under Apache 2.0, with Docker self-hosting documented and a Helm chart for enterprise deployments available on request8. That is a legitimate bridge if you want zero change for now. Treat it as a bridge: a maintenance-mode codebase will receive fewer new features over time, and you take on its operation.
A BYOK control plane: INFRO
INFRO is a hosted gateway on provider accounts you connect, so your providers bill you directly and INFRO charges a subscription with 0% markup on model spend. It covers the cost half of Helicone and adds control: each request records an estimated provider cost at list or negotiated rates, attributed by project, key, model, route and end customer; budgets at organization, project, member and key level refuse a request before the provider is called; routes fail over across your own connections; and an audit log records who changed what. It does not replace Helicone's evaluation and prompt management — INFRO has no prompt management — so teams that relied on those pair it with Langfuse, or choose Langfuse alone.
A migration plan that keeps your history
- Inventory. List every service that calls Helicone, the base URL or headers each uses, and the custom properties and user ids you send. Those properties are your attribution model; they must survive the move.
- Export first. Pull request and cost history out of Helicone into your own warehouse before anything changes. Most tools cannot import another tool's traces, so your history lives where you put it.
- Map the fields. For each Helicone property, find the destination's equivalent: a Langfuse trace attribute, a LiteLLM tag, an INFRO
metadatafield orX-INFRO-*header. - Run in parallel. Send a share of traffic, or a copy of traces, through the new path for a week. Compare daily request counts and estimated cost per day against Helicone; differences point at missing services or mismatched rates.
- Switch the base URL. Change one service at a time behind a flag, starting with the least critical. Changing the base URL in OpenAI SDKs covers each client.
- Keep read access. Keep your Helicone account readable until the new tool has a full billing cycle of history.
What to verify in any destination
- Does it sit in the request path (and can it enforce limits), or beside it (and only record)?
- Whose provider accounts carry the traffic, and what happens when your key fails?
- Does it price requests at your negotiated rates, and say "unpriced" rather than $0 when it has no rate?
- Can you export your data again, so the next migration is easier than this one?
Several products in this market changed owners in 2026; who owns your AI gateway lists them. The durable protection is architectural: your provider accounts in your own name, a standard API shape, and history exported to storage you control.
Summary
Helicone users have time. Use it to name the job Helicone does for you, export the history, and move to the tool built for that job — Langfuse for observability and evaluation, LiteLLM or Bifrost for a self-hosted gateway, or a BYOK control plane for budgets and attribution on your own provider accounts — running in parallel before switching the base URL.