⚠️ Warning: osmosis-ai is still in active development. APIs may change between versions.
Python SDK and CLI for Osmosis AI, a platform for training LLMs with reinforcement learning. Implement an AgentWorkflow and a concrete Grader in Python, then use the CLI to submit evaluation and training runs from an Osmosis workspace directory.
Requires Python 3.12+.
pip install osmosis-ai # CLI + framework-neutral rollout core
pip install "osmosis-ai[server]" # + generic FastAPI rollout server
pip install "osmosis-ai[strands]" # + Strands integration
pip install "osmosis-ai[openai-agents]" # + OpenAI Agents integration
pip install "osmosis-ai[harbor]" # + Harbor backend (uses an externally provided SkyPilot runtime)
pip install "osmosis-ai[rubric]" # + LLM-as-judge rubric evaluation
pip install "osmosis-ai[parquet]" # + Parquet dataset support
pip install "osmosis-ai[full]" # every optional feature
# or with uv: uv add osmosis-aiThere is one distribution, osmosis-ai. The harbor extra installs plain Harbor only: Daytona is retired, and Harbor's skypilot extra must not be installed because the rollout runtime provides SkyPilot. See Installation for product setup and CONTRIBUTING.md for development setup.
Guides, quickstart, and the full CLI reference live at docs.osmosis.ai.
- Quickstart — run the multiply example end to end, from onboarding to evaluation run to training run
- CLI command reference — every
osmosiscommand and flag, plus the--json/--plainoutput contract for AI agents and CI/CD - Workspace setup — repository layout, config files, and Git Sync
- Rollouts — AgentWorkflow, Grader, integrations, and execution backends
- Releases — version history and breaking changes between releases
Building on or contributing to the SDK itself? See the code-anchored developer docs in docs/ (start with docs/architecture.md) alongside CONTRIBUTING.md.
See CONTRIBUTING.md for development setup, testing, linting, and PR guidelines.
MIT License - see LICENSE file for details.