Run a model on your machine

One runtime for many models

Each model family ships its own inference code, so running five models can mean installing five runtimes. OpenDXP, the Open Decision Exchange Protocol, is an open standard that removes that step. It turns a model into a package that any OpenDXP runtime can run, with no code written for that model.

A package is a folder, odxp/ in a model's version, that holds:

• odxp.json, the manifest: the profile, and every file with its SHA-256
• the weights: ONNX for encoder-markers, GGUF for causal-letters
• template.json or prompt.json: how the input is laid out, as data
• calibration.json: the temperatures from lesson 3
• conformance.jsonl: requests, and the answers the model's own code gives

Nothing in a package is executed. Templates are filled in, and the weights run through ONNX Runtime or llama.cpp.

Run a model with one command

systemone run opendxp supersonic-labs/julia-1 \
    --state "Customer: I was charged twice and want my money back." \
    --questions '{"intent": {"type": "choice", "instructions": "What does the customer want?", "criteria": ["refund", "track delivery", "cancel order"]}}'

This answers with Julia 1 on your own machine. The first run takes a few minutes: it downloads the model, sets up a separate Python environment for the runtime, and builds a package for the model if it has none yet. Later runs reuse all of that.

Serve a package over HTTP

The opendxp Python package is the reference runtime, and needs Python 3.11 or newer. Given a package folder, it answers on your machine:

pip install "opendxp[onnx]"     # "opendxp[gguf]" for causal-letters packages
opendxp serve odxp/             # POST /v1/systemone on 127.0.0.1:8790

It takes the same request body as the inference API from lesson 4, at http://127.0.0.1:8790/v1/systemone.

Give it to an AI agent

opendxp mcp odxp/

This serves the package over the Model Context Protocol (MCP), which AI agents such as Claude and Cursor use to call tools. The agent can then ask the model a question and get a probability for every option.

How you know it is the same model

conformance.jsonl records what the model's own code answers on a fixed set of requests. A runtime that gives the same decisions, with every probability within 0.01, runs the model faithfully. A model whose package passes this check shows OpenDXP compatible on its page. The next lesson runs that check yourself, measures the package, and fits its confidence to your data.