Helicone Alternatives: Where Things Stand and How to Keep Your Data

Helicone has been in maintenance mode since Mintlify announced it was acquiring it in March 2026. Its website now says it's no longer part of Mintlify, with details to come. Nothing has been announced about the service closing, but if you rely on it, it's a sensible time to keep a copy of your data and know your options. This guide covers what's known, how to export, and where to go for logging and for testing prompts.

Want to test prompts in a browser? Bring a Helicone dataset into a Prompt test, or paste a prompt and we'll draft the cases. Test your prompt →

What's happened, and when

Sources: Mintlify's announcement, dated 3 March 2026; Helicone's post “Helicone is joining Mintlify”, also dated 3 March 2026, in the Internet Archive's copy from that day; the notice on helicone.ai and the Helicone repository on GitHub, both checked 8 November 2026.

In March, Helicone's founders wrote that the service would “remain live for the foreseeable future”, and offered help to anyone migrating. We don't know what the new notice means for the service, and we haven't guessed. If you're a customer, Helicone is the place to ask.

1. Keep a copy of your data

Whatever happens next, a recent export costs little and keeps your options open.

• Your logs. Helicone's export tool is an npx command, npx @helicone/export, that pulls logs from its API as JSON, JSONL or CSV. You can filter by date range and property, and add --include-body to get the full prompts and completions.
• Your datasets. Helicone's Datasets docs describe two downloads: fine-tuning JSONL (one {"messages": [...]} conversation per line) and CSV for spreadsheets. You can also read a dataset through the API.
• Your prompts. If you used Helicone's prompt management, copy out each prompt's text and the version you run in production.
• Your scores. If you sent evaluation scores to Helicone, note what each score measured and how it was produced, so you can rebuild it elsewhere.

Logs often hold customer data. Store the export as carefully as you stored the logs.

2. Choose where to go next

Helicone does several jobs: a gateway that routes requests, logging and cost tracking, prompt management and datasets. You may need different tools for different jobs.

• Run Helicone yourself. Helicone is open source under the Apache 2.0 licence, and its self-hosting docs cover Docker Compose, Kubernetes and a manual install. This keeps your setup the same, but you take on running it.
• Langfuse is an open-source platform for tracing, prompt management, datasets and experiments, with LLM judges among its evaluators. Its core features are MIT-licensed, and you can self-host it or use its cloud service. It covers logging and testing in one place.
• Promptfoo is an open-source command-line tool for testing LLM apps. You write a config file listing prompts, providers and tests with assertions, and run it from a terminal or in CI. It's for tests, not production logging.
• Your own logging. If you mostly used Helicone for cost and latency, a few fields written to your own database or analytics tool may be enough.
• A browser tool for testing. Spring Prompt's Prompt test is one: paste a prompt, and it drafts test cases and checks for you, runs the models you pick side by side, and shows every answer with its cost. It runs models from many developers. It isn't a gateway and doesn't keep production logs, so it replaces the testing side only.

3. Bring a Helicone dataset into a Prompt test

A dataset downloaded as fine-tuning JSONL uses the same {"messages": [...]} format as chat transcripts, which load straight into a Prompt test with no model call. In a new Prompt test, choose Import from my app, then paste or upload the file.

We don't import Helicone's request log exports as such. For those, pick a few real requests and paste them one at a time, or put them in a dataset first. The other formats we read, including OpenAI Evals files and Promptfoo configs, are in the import table in our OpenAI Evals guide.

A Prompt test takes up to 12 cases and 10 checks, so a large file is cut to the first ones, and the import tells you. Choose the requests that went wrong or nearly did, not twelve easy ones. API keys in the file are removed before anything is stored.

4. Keep a baseline

Before you change anything, save a sample of recent requests with their answers, costs and response times. If you later move gateways, models or prompts, you can check the new setup against what you had, one change at a time.

Bring your cases, or start from a prompt. Run your current model and its alternatives on the same inputs and read every answer side by side. Test your prompt →

Quick answers