Tracer Cloud in closed beta
The hosted platform opened at app.tracerml.ai. Sign up stays open, access is invite-only while we ramp. Bring your own keys, with zero markup on provider tokens.
Changelog
What we shipped.
Dated and public, newest first. We ship fast and in the open across the hosted platform, the routing engine, the open-source library, and the research behind it. Here is the trail.
Partition map and per-cell decision cards
Every tracer renders its full partition. Each cell shows its dominant decision, real example queries, the share of traffic it covers, and its certified accuracy bound, so the routing logic is legible end to end.
Live-query tester
Paste any query and see which cell it lands in and whether it routes to the cheap model or defers to the teacher, before you move a single unit of real traffic.
Certified one-click activation
Each tracer carries a certificate: certified coverage and teacher agreement on held-out traffic. Promote to live routing in one click once it clears your target, and switch back just as fast.
Multi-tier model menu, bring your own keys
Route across a tier menu: the ML surrogate for the predictable slice and your choice of frontier models for the rest. Bring keys for OpenAI, Anthropic, Bedrock, and more, and switch providers anytime.
OpenAI-compatible API and Hermes agent plugin
Point any stack at a drop-in OpenAI-compatible endpoint, or wire tool selection straight into the Hermes agent framework with the native plugin. No prompt changes.
Durable trace storage and live analytics
Every routed query writes its full request and response to durable storage, with a per-tracer redaction toggle. The analytics read live traffic, with coverage, savings, and agreement tracked continuously.
A higher bar for contributions
The open-source repo now asks for evidence over intuition. Model additions and core-parameter changes need a reproducible benchmark on the coverage-vs-agreement frontier. Added a Code of Conduct, CODEOWNERS, and structured issue forms.
Cold-start guide and Router tests
Merged a practical cold-start guide for teams evaluating with small trace sets, plus direct unit coverage for the production Router. Both from the community.
Case studies, with real numbers
Published the first production write-ups. Obside replaced a frontier call per news-by-intent pair with a 38-cell surrogate at 95% saved. getclaw routed agent tool selection through a TRACER classifier and cut end-to-end cost around 50%.
tracer-llm on PyPI
The open-source library shipped: fit a routing policy from your traces, inspect the partition, serve predictions locally. Logistic and gradient heads, calibrated accept gates, an HTML audit report.
The routing method, published and featured
The conformal routing method behind TRACER is written up and featured on Hugging Face Papers. Read it on Hugging Face or arXiv.