Command Line

Reference. cli.py is a plain MCP client: scripts and benchmarks go through exactly the same path as Claude Code. It needs uv. The token comes from $AGENTBOX_MCP_TOKEN or .env; the URL comes from $AGENTBOX_URL (default http://127.0.0.1:8765/mcp).

uv run --script cli.py models
uv run --script cli.py delegate --repo PATH --task-file FILE|- [--model ID] [--venv PATH]
                                [--timeout-min 30] [--max-turns 60] [--no-guide] [--raw]
uv run --script cli.py check    --repo PATH --branch REF --command CMD [--venv PATH] [--timeout-min 15]
uv run --script cli.py cleanup  --repo PATH --branch agent/<id>

• --raw sends the task verbatim, without the subagent preamble. Use it when a benchmark brings its own prompt.
• --no-guide disables the executor guide, for A/B tests.
• Output is the tool's JSON result.

Scripts:

Script	Purpose
scripts/setup.sh	Write .env, secrets.env and config/models.yaml if missing; create data dirs (0700)
scripts/smoke-test.sh [model]	End-to-end check of a model plus the sandbox properties, including run_checks
skills/local-agent-delegation/scripts/make_venv.sh REPO [pkgs…]	Build a venv for python_env inside the agent image
tests/test_gateway_bridge.py	Offline tests for gateway and llm-bridge (no GPU, no Docker)
docs/diagrams/build.py, render.sh	Regenerate and render the Excalidraw diagrams