Osprey Framework

🎉 Latest Release: v2026.5.0 - CalVer adoption, doc-executability CI gate, deploy skill & friction logging

🚧 Early Access Release This is an early access version of the Osprey Framework. While the core functionality is stable and ready for experimentation, documentation and APIs may still evolve. We welcome feedback and contributions!

A production-ready framework for deploying agentic AI in large-scale, safety-critical control system environments—particle accelerators, fusion experiments, beamlines, and complex scientific facilities.

📄 Research This work was presented as a contributed oral presentation at ICALEPCS'25 and will be featured at the Machine Learning and the Physical Sciences Workshop at NeurIPS 2025.

🚀 Quick Start

# Install the framework as a standalone CLI tool (using uv, recommended)
uv tool install osprey-framework

# Create a minimal project to verify your setup
osprey build quickstart --preset hello-world
cd quickstart

# If API keys aren't already in your environment, copy and edit .env:
# cp .env.example .env

# Start a Claude Code agent session
claude

For a production project tailored to your detector, beamline, or accelerator subsystem, install the guided osprey-build-interview skill and run it from Claude Code:

# Install the /osprey-build-interview skill into ~/.claude/skills/
uv run osprey skills install osprey-build-interview

Then start Claude Code in an empty directory and type /osprey-build-interview. The skill walks you through a guided conversation, produces a build profile, and osprey build profile.yml generates a ready-to-use project.

📚 Documentation

📖 Read the Full Documentation →

🧪 Testing

# Run unit tests (fast, no API keys required)
pytest tests/ --ignore=tests/e2e -v

# Run e2e tests (slow, requires API keys)
# ⚠️ IMPORTANT: Use 'pytest tests/e2e/' NOT 'pytest -m e2e'
pytest tests/e2e/ -v

See tests/e2e/README.md and the Contributing Guide for details.

Key Features

• Dual-Mode Orchestration - Plan-first (complete upfront plans) and reactive (ReAct, step-by-step) execution with explicit dependencies and operator oversight
• Control-System Safety - Pattern detection, PV boundary checking, and mandatory approval for hardware writes
• Protocol-Agnostic Integration - Seamless connection to EPICS, LabVIEW, Tango, and mock environments
• Scalable Capability Management - Dynamic classification prevents prompt explosion as toolsets grow
• Production-Proven - Deployed at major facilities including LBNL's Advanced Light Source accelerator

---

📖 Citation

If you use the Osprey Framework in your research or projects, please cite our paper:

@article{10.1063/5.0306302,
      author = {Hellert, Thorsten and Montenegro, João and Sulc, Antonin},
      title = {Osprey: Production-ready agentic AI for safety-critical control systems},
      journal = {APL Machine Learning},
      volume = {4},
      number = {1},
      pages = {016103},
      year = {2026},
      month = {02},
      doi = {10.1063/5.0306302},
      url = {https://doi.org/10.1063/5.0306302},
}

---

For detailed installation instructions, tutorials, and API reference, please visit our complete documentation.

---

Copyright Notice

Osprey Framework Copyright (c) 2025, The Regents of the University of California, through Lawrence Berkeley National Laboratory (subject to receipt of any required approvals from the U.S. Dept. of Energy). All rights reserved.

If you have questions about your rights to use or distribute this software, please contact Berkeley Lab's Intellectual Property Office at IPO@lbl.gov.

NOTICE. This Software was developed under funding from the U.S. Department of Energy and the U.S. Government consequently retains certain rights. As such, the U.S. Government has been granted for itself and others acting on its behalf a paid-up, nonexclusive, irrevocable, worldwide license in the Software to reproduce, distribute copies to the public, prepare derivative works, and perform publicly and display publicly, and to permit others to do so.

---