Installation
Install DeepGym from PyPI with pip or uv, add the extras for the sandboxes and integrations you use, and check that the CLI runs.
DeepGym needs Python 3.10 or newer. The base package depends only on pydantic and httpx (plus tomli on Python 3.10), and runs solutions locally without any sandbox account.
pip install deepgymuv add deepgymExtras
Each sandbox provider and integration has its own extra, so you install only the SDKs you use. Docker and Apple container need no extra: DeepGym drives their command-line tools.
| Extra | Installs | Needed for |
|---|---|---|
daytona |
Daytona SDK | mode='daytona', including desktop (computer-use) environments |
modal |
Modal SDK | mode='modal' |
e2b |
E2B SDK | mode='e2b' |
server |
FastAPI and uvicorn | deepgym serve, deepgym console and the OpenRLHF reward router |
hf |
datasets and huggingface_hub |
Sharing environments on the Hugging Face Hub and dataset importers |
lm_eval |
lm-evaluation-harness 0.4.9+ | lm-eval tasks |
traces |
pyarrow | Parquet trace export |
all |
Everything above except lm_eval |
pip install "deepgym[daytona,server]"uv add "deepgym[daytona,server]"Install from main
These docs follow the main branch on GitHub. To use changes that aren’t in a PyPI release yet (listed under Unreleased), install from there:
pip install "deepgym @ git+https://github.com/DeepGym/deepgym.git#subdirectory=deepgym"uv add "deepgym @ git+https://github.com/DeepGym/deepgym.git#subdirectory=deepgym"A GitHub install doesn’t include the prebuilt console UI, so deepgym console serves only its API and prints how to build the UI.
Check the install
deepgym --helpThe command lists the subcommands: run, run-batch, eval, new, check, qa, harbor, traces, serve, console and the rest. See the CLI reference for each one.