/v1/chat/completions request does not carry, so they do not
fire on LLM traffic today. Gateway-level enforcement is opt-in (observe-first by
default), and an ML ensemble is on the roadmap.
You point traffic at us, get low-overhead gateway routing, full-fidelity
trace capture, tamper-evident audit logs, and merge-gate evals. Premium
positioning, Apache 2.0, no SaaS-blocked license games.
60-second quickstart
Drop two env vars in, get your first trace.
Architecture
Five components, one monorepo, OTel under the hood.
Agent-tool safety
MCP tool-definition-drift + schema validation at the gateway. Each rail backed by a green eval.
205 providers
BYOK across Anthropic, OpenAI, Google, Bedrock, Azure, Cohere + any OpenAI-compatible endpoint.
Why Tracelane
Apache 2.0, full stop
Not GPL-3.0 viral. Not ELv2 SaaS-blocked. Run it however you want.
Caught at the gateway, not after
Tool-definition-drift is flagged or halted inline at the gateway, not
reconstructed from an incident write-up.
Every claim carries an eval
20 conformance evals — 10 fault-tolerance chaos scenarios, 7 gateway-correctness,
and one each for ingest-schema, PII-redaction and prompt-injection. They run in
the canonical repository’s CI; in this public mirror they ship as readable
reference, because six of them assert against fixtures the mirror deliberately
withholds. Structural assertions run green; behavioral ones report as
skipped without a live gateway rather than as a pass.
Tamper-evident ledger
SHA-256 hash chain and Ed25519 per-tenant signing, with byte-identical
verifiers in Rust + Python + TypeScript. Anchored batches carry a resolved
Sigstore Rekor v2 inclusion proof (best-effort, opt-in); universal per-run
anchoring is on the roadmap.
Where to next
- New here? Start with the 60-second quickstart.
- Operating it yourself? Read Onboarding for the self-host checklist.
- Migrating from Helicone, LiteLLM, or Langfuse? See Migrations.
- Auditing the design choices? Browse Decisions.