Skip to main content
Helicone is mostly a proxy with config-via-env + per-request headers. The canonical Tracelane migration path is one CLI command: There are two commands and they do different things.
tlane migrate helicone reads .env, .env.local, .env.production and .env.development, translates each known HELICONE_* var, and rewrites those files and matching source in place. It does not emit .env.tracelane. tlane import-helicone is the non-destructive path: it reads --env <path> and writes a .env.tracelane stanza. --with-config adds a fixed two-provider starter scaffold — it is a template, not derived from your code. --src <dir> only names a directory for you to review by hand; nothing is scanned.

​
Historical requests

Use tlane import-helicone --traces helicone-history.json --dry-run to preview a saved request export, then omit --dry-run to send it. The local cursor makes the import resumable. See Import Helicone history for the input format, counts, captured fields and retention behavior.

​
Env var mapping

​
What you keep

  • Existing OpenAI/Anthropic SDK code — no swap required
  • Per-request end-user attribution — send X-Tracelane-User-Id, or just set OpenAI’s own user field on the request body and Tracelane reads it. It lands on the span as user_id, and the /sessions page shows a User column you can click through to that person’s traces
  • Session grouping — X-Conversation-Id

​
What changes

  • The base URL: https://gateway.tracelane.dev/v1 (or your self-host endpoint)
  • Custom properties have no equivalent. Helicone’s Helicone-Property-* headers become arbitrary span dimensions; Tracelane has no such header. The nearest thing is X-Business-Reference, a single free-form identifier (a case or transaction id) — one value, not a bag of key/values. If you rely on multiple custom properties, tell us; it is not built
  • Response caching has request controls within operator bounds. Send x-tracelane-cache: bypass on any plan to skip lookup and storage. Eligible paid plans can send use or ttl=<hours>; the effective reuse age is bounded by the plan and operator. The response reports x-tracelane-cache-ttl-hours and x-tracelane-cache-bound. Only buffered chat completions are supported; explicit opt-in on streaming chat, embeddings or native messages is refused. Inspect the active configuration at /settings/gateway.
Send an opaque id, not an email address. The user id is stored on the span, and PII redaction rewrites anything email-, phone- or card-shaped before it is written — so [email protected] is stored as [REDACTED:email] and the User column reads “redacted” for every such trace. A UUID or a hash of your internal user id is what you want, which is also what OpenAI and Anthropic both ask for on their equivalent fields.
  • Predictive guardrails inline at the gateway. We found no equivalent inline guardrail layer documented by Helicone as of 2026-08-26.

​
License posture

Helicone’s ai-gateway is licensed GPL-3.0; Tracelane is Apache 2.0 across the full stack. Tracelane is an independent implementation. Licences verified as of 2026-08-26 from each repository’s LICENSE file.
  • Quickstart
  • Providers — the 191 we route to (6 native adapters + any OpenAI-compatible)
  • CLI — full tlane command list