Installation

Installation

Requirements

• Linux, Docker Engine ≥ 27, Docker Compose v2 with gpus: support
• NVIDIA GPU + NVIDIA Container Toolkit (docker run --gpus all must work)
• uv (for cli.py), Python 3 (for setup.sh)
• Claude Code as the planner
• Disk: about 1 GB of images, plus model weights (Nemotron-3.5-Lightning via Ollama is ~25 GB)

Steps

git clone https://github.com/Adithya-Rajendran/agentbox && cd agentbox
scripts/setup.sh
docker compose --profile build build
docker compose up -d
scripts/smoke-test.sh

scripts/setup.sh:

• writes .env with your uid/gid, the docker group id, data paths and a fresh MCP token;
• copies config/models.example.yaml to config/models.yaml;
• creates the data directories as you (Docker would otherwise create missing bind sources as root);
• warns if the agent uid (default 64000) exists on the host.

Override paths at setup time: AGENTBOX_DATA=/data/agentbox AGENTBOX_REPOS_ROOT=~/code OLLAMA_MODELS_DIR=/data/ollama scripts/setup.sh.

scripts/smoke-test.sh [model]:

• creates a throwaway repo under AGENTBOX_REPOS_ROOT;
• delegates a one-line bug fix (default model: the Ollama one; the first run pulls it);
• verifies the diff plus 7 sandbox properties of the live container, then cleans up.

Run it against a vLLM model too: scripts/smoke-test.sh nemotron-3.5-bolt.

Register with Claude Code

claude mcp add --transport http local-agents http://127.0.0.1:8765/mcp --scope user \
  --header "Authorization: Bearer $(grep ^AGENTBOX_MCP_TOKEN= .env | cut -d= -f2)"
ln -s "$PWD/skills/local-agent-delegation" ~/.claude/skills/

Tasks can run for 30+ minutes. Raise the MCP tool timeout in the environment Claude Code runs in: export MCP_TOOL_TIMEOUT=7200000.

Upgrading

git pull && docker compose --profile build build && docker compose up -d

config/models.yaml and .env are yours and are never overwritten.