Comparisons
How the Lightcone Stack relates to other tools for AI-assisted research.
Research is shared today as a paper, sometimes with a code repository, and neither records every choice behind a result. Lightcone Research starts from the idea that agent-driven research should also produce a record of the analysis: its inputs, outputs, decisions and evidence, complete enough for someone else to reproduce and extend it.
The Lightcone Stack keeps that record in astra.yaml, written in ASTRA, an
open specification. You can review, reproduce and cite the analysis as a piece
of work in its own right. This page compares the stack with other tools for AI-assisted research.
At a glance
| Tool | Main focus |
|---|---|
| Lightcone Lab | Deep scientific provenance for AI-driven research, in a JupyterLab workbench |
| Claude Science, Open Science | Research workbenches that track the provenance of artifacts |
| Kosmos, The AI Scientist, Denario | Autonomous research, from question to report |
| Snakemake, Nextflow | Defining and running reproducible pipelines |
Provenance and decisions
Research workbenches such as Claude Science and Open Science save versioned figures and other files, with the conversation, code and logs that produced them. That tells you how each file was made. ASTRA also records why: the scientific decisions behind the analysis, the options considered for each, and the literature that supports them.
Workflow managers
Snakemake and Nextflow are the established way to define and run reproducible pipelines. They capture how each output is computed, but not the scientific reasoning behind the pipeline, and they weren't designed with agents in mind. The Lightcone Stack is built for agents and records the decisions and evidence behind each step. You can use the two together, as the next section shows.
Bring your own workspace
Claude Science comes with its own agent and workspace. With the Lightcone Stack, you keep the agent and the tools you already use.
The stack covers an analysis from start to finish, but its tools are independent, and none requires the
others. You can record an analysis in astra.yaml and run it with your own scripts or workflow manager
instead of the Lightcone CLI. You could also ask the agent in Claude Science to work from astra.yaml.
Researchers stay in the loop
The Lightcone Stack is for researchers who want to understand every step of their analysis and stay responsible for it. Autonomous AI scientists such as Kosmos and Sakana's AI Scientist pursue a research question end to end on their own. Denario can do the same, from generating the idea to drafting and reviewing the paper, though its agents can also run a single step. You receive their conclusions without having made the choices behind them.
Share the whole project
A Lightcone project is one git repository, holding the specification, the code, the data, the results with their run records, and the report. You share it the way you share any repository, with no export step. Git carries the code and the record to any git host, and git-annex keeps the data and results wherever you choose to store them, from a lab server to cloud storage, recording where each file lives.
A collaborator who clones the project can check how every result was made before downloading any data,
and git annex get fetches the files when they need them. When you publish, the agent adds a license
and an RO-Crate description, so the project can be archived
and cited as it stands. The other tools on this page focus on producing results, and leave sharing the
work behind them to you.
Owned by the research community
The Lightcone Stack is open source. We believe the academic community should own, maintain and extend the tools it relies on. Claude Science and Kosmos, by contrast, are proprietary and require a subscription.