OpenJevX
An open, local decision model server for jevx. It answers yes/no, pick-one and rating questions about a JSON state — no cloud call, no API key.
One step
npx git+https://github.com/muthuishere/openjevx.git
openjevxManual — macOS
curl -L -O https://github.com/muthuishere/openjevx/releases/download/v0.5.11/openjevx-darwin-arm64.tar
curl -L -O https://github.com/muthuishere/openjevx/releases/download/v0.5.11/openjevx-model-0.5.2.tar.gz
tar -xf openjevx-darwin-arm64.tar && tar -xzf openjevx-model-0.5.2.tar.gz && ./openjevxManual — Linux
curl -L -O https://github.com/muthuishere/openjevx/releases/download/v0.5.11/openjevx-linux-amd64.tar
curl -L -O https://github.com/muthuishere/openjevx/releases/download/v0.5.11/openjevx-model-0.5.2.tar.gz
tar -xf openjevx-linux-amd64.tar && tar -xzf openjevx-model-0.5.2.tar.gz && ./openjevxLinux on ARM (arm64): use openjevx-linux-arm64.tar instead.
Windows: download openjevx-windows-amd64.zip and openjevx-model-0.5.2.tar.gz, unpack both into the same folder (tar -xzf openjevx-model-0.5.2.tar.gz) and run openjevx.exe.
Docker
docker run -d -p 127.0.0.1:21160:21160 \
-e OPENJEVX_PASSWORD=<12+ characters> -e OPENJEVX_API_KEY=<16+ characters> \
ghcr.io/deemwar-products/openjevx:v0.5.9The prebuilt image is for linux/amd64 and linux/arm64. To build it yourself:
git clone https://github.com/muthuishere/openjevx.git && cd openjevx
OPENJEVX_PASSWORD=<12+ characters> OPENJEVX_API_KEY=<16+ characters> docker compose up -d --buildTalk to it from jevx
jevx profile add openjevx http://127.0.0.1:21160/v1/systemone --model openjevx
jevx profile use openjevx
jevx is "Is this urgent?" < email.txtWhat it is
Default port 21160, endpoint /v1/systemone. device is auto, cpu, or gpu in openjevx.json.
Dashboard
Open http://127.0.0.1:21160/ while it runs: request counts, latency p50/p95/p99, errors, recent requests. No default password: set "password", or read the one the server generates into openjevx.password.
License
Apache-2.0. Credits in CREDITS.
Read the journey → how it went from a 77.4% fine-tune to a shipped model that's 98.1% on held-out software-role decisions, and where it's still weak.