Installation¶
Requirements¶
- Python 3.11, 3.12, or 3.13.
- Windows, Linux, or macOS for Guard, MCP, Replay, and Benchmark.
- Docker on Linux or through WSL2/Docker Desktop for Sandbox.
- Node.js only when developing or packaging the VS Code extension.
Install the current alpha¶
Install the public alpha from PyPI. The distribution name is agent-loop-guard-runtime, and the installed command is alg.
The package is published through a PyPI Trusted Publisher bound to this repository's publish-pypi.yml workflow. No long-lived PyPI token is stored in GitHub.
Optional dependency groups¶
The default installation includes Guard, MCP, Replay, the web UI, and JSONL benchmarks. Install optional groups from a cloned checkout:
| Extra | Install command | Adds |
|---|---|---|
bench |
pip install -e ".[bench]" |
PyArrow, DuckDB, MLflow |
sandbox |
pip install -e ".[sandbox]" |
Marker extra; Docker remains external |
docs |
pip install -e ".[docs]" |
MkDocs Material |
dev |
pip install -e ".[dev]" |
pytest, coverage, Ruff, HTTP mocks |
mcp is currently a compatibility marker and adds no package beyond the default dependencies.
Verify the installation¶
doctor checks Python, storage, the configured port, Docker, and WSL. Missing Docker is a warning because only the Sandbox module depends on it.
Upgrade¶
Upgrade the public package with the same tool used for installation:
Database migrations run automatically at application startup.
Uninstall¶
Remove the Python package with the tool manager that installed it. Local project state remains in agent-loop-guard.yml, .agent-loop-guard/, and the configured SQLite path until removed manually.