Install OpenJet with your coding agent

Inspect the machine

Check OS, architecture, Python (3.10 or newer), free disk, RAM, existing OpenJet installations, and available package tools. On NVIDIA hardware run:

nvidia-smi --query-gpu=index,name,memory.total,memory.free --format=csv,noheader,nounits

On Linux inspect RAM with free -m and disk with df -h. On macOS use uname -m, sysctl -n hw.memsize, and df -h; Apple silicon shares RAM between CPU and GPU. On Windows use PowerShell Get-CimInstance Win32_ComputerSystem and Get-PSDrive -PSProvider FileSystem, plus nvidia-smi when available. A Jetson or AMD machine may have GPU acceleration without nvidia-smi; use OpenJet's hardware detection below.

Do not add GPU VRAM together when choosing a model. Check available memory as well as total memory. Do not change drivers, delete models, or install privileged dependencies without the environment's required authorization.

Install

Reuse an installation if openjet setup --help lists --non-interactive. Otherwise install the package in isolation:

pipx install open-jet

Use the actual launcher path reported by pipx if the current shell's PATH does not include it. If the published package does not yet contain --non-interactive, install from the source checkout instead; do not fall back to the wizard.

For a source installation on Linux or macOS:

git clone https://github.com/l-forster/open-jet.git "$HOME/.openjet/source"
cd "$HOME/.openjet/source"
./install.sh

If the destination already exists, inspect and reuse a valid checkout rather than overwriting it. Use ./install-macos.sh on macOS if preferred. The installer exits after installing and reports launcher paths. The checkout's .venv/bin/openjet works without a PATH change. On Windows, clone a checkout, run install.bat from it, and use .venv\Scripts\openjet.exe. If pipx is missing, use the source installer rather than installing into system Python.

Inspect the recommendation without changing setup

Use the installed launcher for all commands below:

openjet setup --non-interactive --dry-run --json

This detects hardware and returns the recommended model path, context, device, download size, configuration target, and available catalog IDs. Check these against the machine inspection. OpenJet chooses from its own catalog; do not invent model names or maintain a separate model/VRAM table.

Run setup from a directory without a project .openjet/config.yaml overlay. If OpenJet reports a configuration conflict, move to a neutral directory and inspect again. Preserve a compatible existing local model. If the user requested a particular model, use its catalog filename or a local .gguf path with --model.

Configure without the wizard

For the automatic recommendation:

openjet setup --non-interactive --yes --json

For a specific catalog model returned by dry-run:

openjet setup --non-interactive --yes --model "<catalog filename>" --json

For an existing GGUF:

openjet setup --non-interactive --yes --model "/absolute/path/model.gguf" --json

Replace placeholders with actual values. --context-tokens <number> may lower the recommended context. --yes accepts model download and runtime provisioning. The command never opens the wizard or TUI, keeps other model files, tests inference, and saves configuration only after verification succeeds.

JSON is written to stdout; progress and build/download logs go to stderr. Let builds and downloads finish. On a nonzero exit, read stage and error, fix that specific problem, and retry the same command. Missing source-build tools typically require git, CMake, and a compiler; CUDA builds can also require the CUDA toolkit. Use the platform deployment docs for details. Do not silently switch to cloud inference or repeatedly download different models.

Verify and hand off

Require status: "ready" and verification.passed: true. Configuration or a successful download alone does not prove inference works. Then run:

openjet status --json

Run status from the user's intended project directory too; confirm no project overlay changes the selected model. Do not launch the interactive TUI in the installing agent's terminal.

Finish by telling the user their detected hardware, model and quantization, actual backend from verification.backend, context size, and that local inference passed. Include these usage commands, with their project path and the actual executable path where needed:

cd /path/to/your/project
openjet

Explain: enter a coding task and press Enter; use @file to include a file, /help for commands, and /exit to quit. Include a one-shot example:

openjet chat "Explain this repository"

Mention material limitations such as CPU-only performance or a fallback backend. If installation remains blocked, state the failed stage and exact recovery command instead of claiming it is ready.