Lightcone Research

Tools for transparent, reproducible and verifiable science.

lightconeresearch.org

Computational research is full of methodological choices that quietly shape the results: how a dataset is filtered, how a model is parameterized, what cutoff defines an outlier. These choices are often defensible, often debatable, and almost always invisible by the time the paper is written.

We build open tools that make those choices explicit and keep the record for you: your AI agent plans the analysis with you, records each decision with the evidence behind it, and every result comes from a recorded run you can trace back and reproduce months later.

1. Install uv, version 0.12 or later (check with uv --version).

2. Add the Lightcone plugin to your coding agent. It installs the Lightcone CLI it needs on first use.

   Claude Code

   claude plugin marketplace add LightconeResearch/agent-skills
   claude plugin install lightcone@lightcone-research

   Codex (approve the plugin's hooks when Codex asks: they are how the plugin checks the agent's work)

   codex plugin marketplace add LightconeResearch/agent-skills
   codex plugin add lightcone@lightcone-research

   For the desktop apps, see the agent-skills instructions.

Then start your agent in an empty directory and give it your research question:

I want to start a new Lightcone analysis: <your research question>

The words "new Lightcone analysis" are enough for your agent to pick up the Lightcone plugin. It works through the plan with you, records every decision in an astra.yaml, then runs the analysis with full provenance.

👉 Follow the full quick start for the next steps.

These projects are in early alpha and evolve quickly: interfaces may change between releases. Bug reports and feedback are very welcome.

• One scientific specification for your project. astra.yaml is the single source of truth. Every input, output, recipe, and decision lives there.
• Decisions and alternatives. Every meaningful methodological choice is a named decision with options, a default, a rationale, and (where possible) literature evidence behind it.
• Provenance you can prove. Every output is committed together with a content-addressed manifest and a re-runnable run record. Tampering and drift are detectable, not hopeful.
• Agents do the typing, you do the science. The agent handles the implementation grind; you make the scientific choices.

• Questions, ideas, or want to try the stack on your own research? Come talk to us on Discord.
• Found a bug? Two ways to report it:
  • Open an issue using our templates.
  • From your agent session, install the Lightcone Feedback plugin and ask: "Use Lightcone Feedback to report this: …". It drafts the issue for your approval and routes it to the right repository.
• Looking for the deeper docs? Everything is on docs.lightconeresearch.org.
• Want to contribute? We welcome contributions across the stack: schema design, the agent surface, the execution engine, the docs, and the rough edges you hit when you use any of this for real work. Check a repository's CONTRIBUTING.md, where it has one, for development setup and conventions. Smaller fixes are welcome as PRs without prior discussion; for larger changes, opening an issue first is appreciated.

Our code ships under BSD 3-Clause unless a repository says otherwise. We aim to keep things permissively licensed so they can be used freely in academic and commercial research alike.