Installation

Requirements

OS	Linux only — enroot is Linux-only; Debian/Ubuntu and RHEL/CentOS/Fedora are both supported (install commands below).
Python	3.12.x — pyproject.toml enforces >=3.12,<3.13. Newer or older versions will fail to install.
GPU	NVIDIA GPU(s) with CUDA. Required by vLLM and the verl trainer. The shipped train_*.sh scripts target 1 to 8 GPUs (trainer.n_gpus_per_node=1 for train_example.sh, up to =8 for the bigger tasks); per-GPU VRAM scales with model size (Qwen 3B fits comfortably on a single 80 GB card; 32B benefits from multi-GPU).
Sudo	Needed once, to install enroot (used by code_interpreter, alfworld, webshop, scienceworld, and any container-backed tool). Skip if you only need the lighter tools (calculator, retrieval API, etc.).
Disk	A few hundred GB headroom recommended: HuggingFace model weights, enroot images under $XDG_CACHE_HOME/AgentFly/enroot/, and (for retrieval tasks) the Wikipedia corpus + FAISS index downloaded into $XDG_CACHE_HOME/AgentFly/data/search/ (~30 GB).
Conda	The reference setup uses conda (bash install.sh assumes it). Other environment managers work but are not the documented path.

If you only want to run agents — no RL training, no container-backed tools — pip install -e . alone is enough. The [verl] extras and the enroot install are training- and environment-specific.

Install With Script

To install dependencies, run the following script in conda environment. We default to use python3.12.

bash install.sh

Step-by-Step Installation

Alternatively, you can customize the installation by following these steps:

1. Clone the repository and initialize submodules:

   git clone https://github.com/Agent-One-Lab/AgentFly
   cd AgentFly
   git submodule init
   git submodule update
   

2. Initialize and install dependencies

   Basic python packages installation:

   pip install -e .
   pip install -e '.[verl]' --no-build-isolation
   

   Some of our tools & environments are managed by enroot backend. To use them, please install enroot (sudo required). Such tools include code_interpreter, retrieval, webshop, alfworld, sciencworld.

   # enroot install
   # Debian-based distributions
   arch=$(dpkg --print-architecture)
   curl -fSsL -O https://github.com/NVIDIA/enroot/releases/download/v3.5.0/enroot_3.5.0-1_${arch}.deb
   curl -fSsL -O https://github.com/NVIDIA/enroot/releases/download/v3.5.0/enroot+caps_3.5.0-1_${arch}.deb # optional
   sudo apt install -y ./*.deb
   
   # RHEL-based distributions
   arch=$(uname -m)
   sudo dnf install -y epel-release # required on some distributions
   sudo dnf install -y https://github.com/NVIDIA/enroot/releases/download/v3.5.0/enroot-3.5.0-1.el8.${arch}.rpm
   sudo dnf install -y https://github.com/NVIDIA/enroot/releases/download/v3.5.0/enroot+caps-3.5.0-1.el8.${arch}.rpm # optional
   

3. Optional

   Search requires redis to cache results, an optional way to install with conda:

   conda install conda-forge::redis-server==7.4.0