model_list to map logical model names to provider
upstreams. tlane import-litellm reads that file and prints the equivalent
Tracelane routing stanza:
--config is the flag the command expects; a bare path argument is not read.
Caller-side code that calls litellm.completion(...) keeps working once you set
its api_base to Tracelane’s OpenAI-compatible /v1/chat/completions endpoint —
you keep the LiteLLM SDK, no swap needed.
Mapping
What the importer actually reads, and the key it emits (packages/cli/src/commands/import-litellm.ts):
router_settings.timeout and router_settings.fallbacks are not read by the
importer — carry those across by hand.
What the gateway gives you structurally
These are properties of Tracelane’s own request path. We make no claim here about any other proxy’s security history — check that against its own advisories.- No endpoint that rewrites gateway config. Model-to-provider routing is
compiled into the binary and fails closed on an unknown model (
400 unroutable_model, no default upstream).tracelane.yamlis read once at process start; no endpoint reloads it at runtime. The/v1/*write routes that do exist (BYOK keys, prompt versions, alert rules) are authenticated, tenant-scoped and ledgered; none of them is global configuration. - No interpreter on the request path. The gateway is a single Rust binary; it
never
evals or imports config as code, because there is nothing to eval it with. .pthsupply-chain scan. CI fails the step on any stray.pthfile in the tree (.github/workflows/ci.yml,secret-scanjob).
.pth scan is a CI step that exits non-zero on a hit.
Related
Competitor details verified as of 2026-08-26 against public documentation and public repositories. Where a capability is marked absent, we found none documented at that date — not that the vendor lacks it. If anything here is inaccurate, please open an issue.