Tracer

Tracer is a Y Combinator backed lab and product company that works on getting more performance per token from AI systems. Its open source library, installed as tracer llm, trains a lightweight surrogate model from an LLM's own production traces. An acceptor gate then serves predictable classification inputs locally and defers uncertain cases to the original model. It is aimed at AI teams paying for repetitive calls to frontier models. The site lists case studies from teams that replaced such calls with surrogates, including one that cut agent cost by roughly half by routing tool selection through a classifier. A decorator called tracer.watch observes calls and sends them to Tracer Cloud, a hosted platform that accumulates data and optimizes once enough is collected, and a JavaScript watcher package offers the same for TypeScript. The company also offers Echo, an ensemble of specialist open models presented as one OpenAI compatible endpoint that allocates the right amount of intelligence to each task and can be customized to a workload. The site also hosts research papers, a blog, documentation and a changelog. The site publishes no prices for Tracer Cloud or Echo, and the route to a commercial conversation is a booked call. The paper reports 83.2 percent local coverage on the Banking77 benchmark at 0.959 teacher agreement.