16 projects · checked 2026-09-20
Run a decision model yourself
Within days of the commercial API shipping, people started reimplementing its one endpoint. 16 of those projects are live enough to be worth listing.
Why this exists at all
The decision API is one endpoint with three primitives. That is a small enough surface that a competent developer can reimplement it in a weekend, which is exactly what happened. The interface spread faster than the models did.
Three different things get called self-hosting here, and they are not interchangeable.
A compatible server gives you the same call shape as the commercial API in front of weights you choose, so the client code does not change. Anadapter wraps a model you already run. A model plus serverships its own weights, which means you are also adopting whatever that model can and cannot do.
None of these have been measured by us. What follows is what each project says it is, with its star count and last push date so you can judge whether anyone is still on it. The models they serve are inthe directory, where 21 of the entries ship in a format you can run locally.
Compatible server
Exposes the same call shape as the commercial API, so client code does not change.
Adapter
Wraps a model you already run, putting the decision interface in front of it.
Model plus server
Ships its own weights rather than serving somebody else’s.
Running the weights directly
If you do not need the API shape, you can load a model straight into a local runtime. Each guide lists the models in the directory that ship in that format.
Run a decision model with llama.cpp
GGUF is the format that made local inference ordinary. For decision models it matters more than usual, because the whole argument for the category is cost per call.
Run a decision model on Apple Silicon
MLX uses unified memory on M-series chips. For a model in the few-hundred-megabyte range this is the fastest local option on a Mac, and it needs no GPU of its own.
Run a decision model with ONNX Runtime
An ONNX export loads without PyTorch. That matters for slim containers, for edge devices, and for running a decision inside a browser.
Run a decision model on device with Core ML
Core ML puts the model on the Neural Engine, so the decision happens on the phone or the laptop and nothing leaves it. The strongest privacy story in the category.
One thing to check before you commit
Several of these projects claim to match or beat the commercial model. None of those claims has been checked by anyone outside the project that made it, and we have not checked them either. Treat a README benchmark as a starting point for your own test, not as a result. The one model we did measure shows what checking it actually involves.
Find out what your data actually supports
We turn your historical decisions into a frozen test set, then tell you which model, which threshold, and how much of it you can safely automate.