GitHub · Serving and API · Agents and tools

bladedevoff/stuntd

Cutting an existing LLM bill without changing the application.

The repository's GitHub card, read on 2026-09-24

What it does

Sits in front of an LLM API, records the typed decisions that pass through it, trains a small head on those records, and answers subsequent calls locally when the head is confident enough.

The repository's own description: “Local proxy that learns your app's typed LLM decisions and answers them with a Laya head. Jev and OpenAI compatible.”

How it works

Recorded decisions become training data for a Laya-compatible head, and the proxy decides per request whether to answer locally or fall through to the model behind it — which is what makes the accuracy question a question about your traffic rather than about the model.

The repository, by the numbers

Stars

Licence

Apache-2.0

Stack

Python · Laya head · local proxy

Last push

2026-09-24

Read from the GitHub API on 2026-09-24. Stars and the last push move daily — quote them with the date, the way we do.

What we checked

the README, read 2026-09-24
the fall-through behaviour and the licence
the stated per-decision latency and the percentage it is claimed to answer

What we did not check. The 90% figure. It depends on the traffic you feed it, the repository says as much, and we have not run it against ours.

Open the repository ↗All builds

The rest of the topic

Jev, the reference →

Start here

Laya, the reference What it is, what we measured running it, and what it is not.
Laya and Jev, compared The two models on the same questions, including the ones Laya loses.

Columns

Built with Laya 18 you are here Ports, servers, agents and games, with what we checked.
What X is saying 22
What people filmed 9
What Hacker News made of it 6
Docs, weights and reading 15

The topic

Every record, by kind of work The same rows cut a second way, across every column.
What self-hosting costs Hardware, memory, and the failure modes of running it yourself.
What the licence allows Apache-2.0 weights, three checkpoints, and the limits that come with them.

By kind of work

Runtime ports 23 Getting the weights onto someone else's hardware: MLX on a Mac, Core ML on the Neural Engine, ggml in C++, candle in Rust, ONNX in Node.
Serving and API 8 Standing the model up as something else can call, usually behind Jev's own wire format.
Agents and tools 6 Typed decisions inside an agent loop: routing, evals, guardrails, memory, and the plumbing around them.
Browser and computer use 2 Deciding the next click, key or field on a real screen.
Games and demos 9 The model playing something you can watch. Most of this model's traffic arrived through one of these.
Benchmarks and evaluation 19 Numbers somebody actually ran — including the ones that say Laya lost.

What it does

How it works

The repository, by the numbers

What we checked

More like this

receptron/laya

wdobry/laya-playground

lkarlslund/laya.cpp