Helicone alternatives after the Mintlify acquisition

Helicone earned its place in your stack with one line of config: point the SDK’s base URL at its gateway, and every call came back logged, priced, and cacheable. On March 3, 2026, Helicone announced it was joining Mintlify.¹ The line still works. What stands behind it has changed: feature development has ended, the founders have moved to Mintlify to build AI knowledge infrastructure, and the platform that served 16,000 organizations (Helicone’s own count, given in the announcement) runs on in maintenance mode.

Mintlify’s side of the announcement commits to security updates, bug fixes, and support for new models. It also promises to “work closely with every customer to support a smooth migration to another platform.”² Offering to help every customer leave is a signal about where the product is headed. What that signal is worth with no date attached is the next section’s question.

Migrate by surface, and let the request path set your urgency. If your traffic runs through Helicone’s proxy, replace the gateway first: self-host LiteLLM if you want to own the path, OpenRouter for hosted multi-provider routing, Cloudflare AI Gateway if your stack already terminates at Cloudflare, Portkey for the closest shape to what Helicone was. If you used async logging, nothing of Helicone’s sits in your hot path, and you are choosing an observability platform on ordinary criteria; start from what separates Langfuse, LangSmith and Braintrust once you read their licenses. Leave a proxy in place, though, and every model call you make keeps running through a product no one is paid to advance.

What the two announcements commit to, and what neither one dates

Both posts went up on March 3, 2026. Helicone’s says its services “will remain live for the foreseeable future in maintenance mode,” and lists what that includes: security updates, new models, bug and performance fixes.¹ Mintlify’s repeats the commitment and adds the migration promise.²

Both commitments are self-refereed: the acquirer and the acquired describing their own obligations to you, with no date attached and nobody outside the deal holding them to anything. “The foreseeable future” cuts both ways. There’s no deadline forcing your hand today, and no floor under the service either.

Contrast the other migration notice this category produced in 2026: hosted OpenAI Evals is winding down on a published schedule, read-only on October 31 and shut down on November 30. A published date lets you plan a cutover, and an undated commitment leaves you planning around intent.

Helicone’s wind-down has no schedule at all.

The migration guides split by date, and one of them has the signup status wrong

Migration guides for this query are easy to find, and the material splits cleanly by when each one was written. Anything published before March 3 describes a roadmap that stopped existing that day.

At least one of the newer guides carries the opposite error, and it is checkable against the source. A roundup published on July 10, 2026 tells its readers that under maintenance mode “the roadmap is frozen, and new signups are disabled entirely.”³ Helicone’s homepage was still offering its sign-up flow and a seven-day free trial, with no credit card required, when we loaded it on August 2, 2026.⁴ The frozen roadmap is accurate, and it is the half of that sentence the two announcements actually support.

Status facts in this category change month to month. Every claim on this page carries the date we read it at source, and that applies to this page as much as to any other: re-check before you commit to a migration.

Whether Helicone sits in your request path decides how fast you move

Helicone’s docs describe two ways to integrate, and name the trade plainly. The proxy is “as easy as changing the base URL,” and because it sits in the request path it carries the gateway features: caching, rate limiting, key management. Async logging keeps the log write off the critical path, which the same page presents as the reason a Helicone outage “will not affect their app.”⁵

Maintenance mode moves that trade onto your risk register. A gateway that stalls takes the request down with it, while a logger that stalls costs you a log line. The concentration risk is the same one that applies to routing every tool through one MCP gateway: whatever sits in the path shares the path’s fate. So proxy users should move first, and async users can choose calmly.

There is a second reason proxy users can’t switch the gateway off and call it done. Removing Helicone’s proxy also removes the cache and the fallback logic riding on it, so a replacement has to cover those jobs as well as the routing. The table below records that coverage candidate by candidate.

Four gateways cover the proxy job, each with a different custodian

They differ most on who operates the path once Helicone leaves it.

LiteLLM puts the gateway in your own infrastructure. The core repository is MIT-licensed with a separate enterprise folder, and the proxy carries the full gateway toolkit: virtual keys, per-key budgets and spend tracking, load balancing, fallbacks, caching, and rate limits.⁶ You stop depending on anyone’s acquisition plans, and you start operating a service.

OpenRouter is the hosted counterpart: one API over the major providers, with provider inference pricing passed through at “no markup.” Two separate charges sit on top of that. Buying credits by card costs 5.5% of the purchase with a $0.80 minimum, and bring-your-own-key traffic is free for the first 1 million requests a month and then costs 5% of what the same call would have cost through OpenRouter.⁷ On the two jobs Helicone’s proxy did beside routing, it answers one: a provider error falls through to the next provider automatically. It runs no response cache of its own, and instead routes to the provider endpoints that support prompt caching, keeping a session on the same one so those provider-side caches stay warm.⁸ It gives you nothing to operate, and it puts a different company in exactly the seat Helicone held. Take that trade deliberately.

Cloudflare AI Gateway absorbs the job into infrastructure you may already run. Its documentation lists analytics, logging, caching, rate limiting, and request retry with model fallback, available on all Cloudflare plans.⁹ If your traffic already terminates at Cloudflare, this is the lowest-friction landing.

Portkey is the nearest thing to a like-for-like: an open-source gateway with a hosted observability platform above it, an OpenAI-compatible URL swap covering more than 250 models by Portkey’s count, and a hosted free tier of 10,000 requests a month. Its gateway documentation lists simple and semantic caching, fallbacks, automatic retries and load balancing, so both of the jobs riding on Helicone’s proxy have somewhere to land.¹⁰ Portkey’s own measurement of its added latency is 20 to 40 ms per call.¹¹

TABLEShow full table (4 rows)Showing full table (4 rows)

Alternative	Who operates the path	What it covers of Helicone’s gateway job	Pick it when
LiteLLM	You (self-hosted; MIT core)	Virtual keys, budgets, caching, rate limits, fallbacks	You want nothing third-party left in your hot path
OpenRouter	OpenRouter (hosted)	Multi-provider routing, automatic failover to the next provider, a BYOK lane; no cache of its own	You want routing without operating anything, and you accept a new custodian in the old seat
Cloudflare AI Gateway	Cloudflare (edge)	Caching, rate limits, retries, model fallback	Your stack already terminates at Cloudflare
Portkey	You or Portkey (open-source gateway; hosted platform)	Simple and semantic caching, fallbacks, retries, load balancing, plus an attached observability layer	You want the closest shape to what Helicone was, in one product

We have not run these four head to head. No latency or throughput number in this section is ours; the one performance figure here is Portkey’s own, quoted as the vendor’s claim and cited at its source.

If Helicone only logged your calls, you are choosing an observability platform

That market is already mapped, and we won’t remap it here. Licensing and data custody decide it, and our three-way on Langfuse, LangSmith, and Braintrust works through both against the vendors’ own documents. One fact transfers directly: Langfuse sits closest to Helicone’s open-source posture, with all product features MIT-licensed and self-hostable without usage limits, and only enterprise governance modules such as SCIM and audit logging held back.¹²

If the live question in your org is whether the enterprise suite you already pay for can absorb this job instead, that is the Datadog question, and our enterprise LLM observability comparison against Langfuse takes it up.

Whichever platform receives your traces, the migration is a chance to drop a habit the old dashboard tolerated: a pass rate with no test behind it. Before you trust a green number in the new tool, check whether the difference you are seeing is statistically significant, and put an interval on any rate you report with the pass-rate confidence interval calculator.

Self-hosting the repo trades a migration for a maintenance obligation

Every option so far moves you off Helicone’s code. There is one that keeps it: run Helicone yourself. The main repository is Apache-2.0, public, and was not archived when we checked on August 2, 2026,¹³ and that license survives whatever Mintlify decides next. The code cannot be taken from you.

The team can. The people who wrote it now build Mintlify’s products, and the maintenance-mode commitments in the announcements describe the hosted service, stated as the vendor’s intent rather than as anyone’s contract. Adopt the repo and you become the maintainer of record for a gateway in your own hot path.

That position suits a team that already self-hosts, reads the codebase, and would fork on any vendor wobble anyway. It does not suit a team reaching for it to avoid choosing, because it converts a one-time migration into an open-ended obligation.

Staying put is defensible on exactly one surface

Granted, nothing in the source record puts a deadline on an async-logging user. Security updates, bug fixes, and new-model support are all committed, signups were still open on August 2, 2026, and there is no shutdown date to beat.¹⁴ Five months after the acquisition, every one of those commitments still stood where we could check it.

The proxy is where that comfort runs out.

So the order of operations we would run:

1. Find which surface you are on. A Helicone base URL in your client or environment config means the proxy lane; an SDK logger or exporter means async.
2. Proxy lane: pick a gateway from the table, stand it up beside Helicone, repoint the base URL, and verify your cache-hit rate and fallback behavior before you tear the old path down. A replacement that quietly drops caching shows up as a cost and latency change rather than as an error, so nothing pages you when it happens.
3. Async lane: choose the observability platform on the licensing and custody axes, on your own schedule.
4. Re-read the two announcements before you commit either way. Every date on this page is an as-of, and this category produced two migration notices in 2026 alone.

Once the logging question is settled, the harder one opens: whether anything you log shows the answers were right. We draw that line between a finished run and a correct one in AI observability vs evals. A migration is a good moment to ask it, because you are already touching every call site.

Footnotes

1. Helicone, “Helicone is joining Mintlify” (March 3, 2026): “Helicone’s services will remain live for the foreseeable future in maintenance mode,” scoped to security updates, new models, and bug and performance fixes; the post also gives the platform’s totals of 14.2 trillion tokens processed, 16,000 organizations, and 33 million end users tracked. https://www.helicone.ai/blog/joining-mintlify (as of 2026-08) ↩ ↩² ↩³

2. Mintlify, “Mintlify acquires Helicone to redefine AI knowledge infrastructure” (March 3, 2026): Helicone continues in maintenance mode, and Mintlify will “work closely with every customer to support a smooth migration to another platform”; Helicone’s gateway and observability technology folds into Mintlify’s assistant, agent, and workflow products. https://www.mintlify.com/blog/mintlify-acquires-helicone (as of 2026-08) ↩ ↩²

3. PostHog, “The best Helicone alternatives & competitors, compared” (published July 10, 2026): describes Helicone under maintenance mode as “the roadmap is frozen, and new signups are disabled entirely.” The roadmap half matches the two announcements; the signup half did not match Helicone’s homepage on the day we read both. https://posthog.com/blog/best-helicone-alternatives (read 2026-08-02) ↩

4. Helicone homepage: an acquisition banner (“Helicone Joins Mintlify”) links the announcement, and the sign-up call to action was still live above the line “No credit card required, 7-day free trial.” https://www.helicone.ai/ (checked 2026-08-02) ↩ ↩²

5. Helicone, “Proxy vs Async Integration” documentation: the proxy integration is “as easy as changing the base URL” and carries the gateway features (caching, rate limiting, key management), while async logging keeps logging off the critical path so that an outage or network issue “will not affect their app.” https://docs.helicone.ai/references/proxy-vs-async (as of 2026-08) ↩

6. LiteLLM (BerriAI) repository and proxy documentation: MIT license with a separate enterprise folder; the proxy server provides virtual keys, spend tracking and budgets, load balancing, fallbacks, caching, and rate limiting, with the docs’ own claim of “100+ LLMs in a unified interface.” https://github.com/BerriAI/litellm and https://docs.litellm.ai/docs/simple_proxy (as of 2026-08) ↩

7. OpenRouter FAQ: a unified API with “no markup on inference pricing”; platform fee of 5.5% ($0.80 minimum) on card credit purchases; bring-your-own-key usage free for the first 1M requests per month, then 5% of what the same model and provider would cost through OpenRouter. The same FAQ states that “If a provider returns an error OpenRouter will automatically fall back to the next provider.” https://openrouter.ai/docs/faq (as of 2026-08) ↩

8. OpenRouter, prompt-caching documentation: caching is provider-side (“you can enable prompt caching on supported providers and models”), with OpenRouter using provider sticky routing to keep a session on an endpoint whose cache is already warm. OpenRouter documents no response cache of its own. https://openrouter.ai/docs/features/prompt-caching (as of 2026-08) ↩

9. Cloudflare, AI Gateway documentation: analytics, logging, caching, rate limiting, and request retry with model fallback across multiple providers, “available on all plans.” https://developers.cloudflare.com/ai-gateway/ (as of 2026-08) ↩

10. Portkey, AI gateway documentation: the gateway’s listed features include “Cache (Simple & Semantic),” “Fallbacks,” “Automatic Retries,” “Load Balancing,” “Conditional Routing,” and “Request Timeout.” https://portkey.ai/docs/product/ai-gateway (as of 2026-08) ↩

11. Portkey documentation: an open-source gateway with a hosted platform, supporting “over 250 AI models,” integrating via an OpenAI-compatible gateway URL; hosted free tier of 10k requests per month; Portkey’s stated added latency is 20 to 40 ms per call. https://portkey.ai/docs/introduction/what-is-portkey (as of 2026-08) ↩

12. Langfuse, open-source overview: “Everything outside the /ee folders is MIT-licensed,” with tracing, evaluations, prompt management, experiments, annotation, and the playground MIT-licensed without usage limits; SCIM, audit logging, and data-retention policies require a commercial license when self-hosting. https://langfuse.com/docs/open-source (as of 2026-08) ↩

13. Helicone GitHub repository: Apache-2.0 license, public, not archived. https://github.com/Helicone/helicone (checked 2026-08-02) ↩