Rare disease
“From a patient’s symptoms and exome, shortlist the likely diagnosis.”
Reproducible analysis your agent can run
Easy, fast, and cost-efficient
science for everyone.
Ask for an analysis in plain language. An AI agent picks the right reproducible tool, shows you the pipeline, and runs it so the result repeats. For research and decision support, not a medical device.
An ever-growing catalog: pick a field or watch them cycle. Items marked ◇ are on the roadmap, and new solutions land with every release.
The agent in action
Watch an agent compose a pipeline
A recorded agent searches the catalogs, inspects a solution, and outlines a reproducible pipeline. Open the playground to run one live against your own backend.
Hit play to watch the agent search the catalogs, inspect a solution, and compose a pipeline. Open the playground to run an agent live against your backend.
This hosted demo can install and run tools, so you can try the full workflow. A local install ships read-only by default (search, inspect, plan, validate); you enable install, run, and deploy yourself, and the agent asks before each. Tools always run on the server that hosts them, never a hidden cloud.
How Album works
Four ideas explain the whole model. Each step an agent takes maps to one of them.
-
01
Catalog
A versioned index of ready-made solutions, like a package registry.
album_list_catalogs · album_search_solutions -
02
Solution
A self-contained tool with declared inputs, outputs, and citations.
album_get_solution_info -
03
Isolated environment
A pinned conda environment, built with micromamba, so runs don't drift.
conda-forge · micromamba -
04
Coordinates
A
group:name:versioncatalog:group:name:versionaddress. The same coordinates rebuild the same environment and give the same result.
Composing a pipeline
A pipeline chains solutions: each step's output feeds the next.
ngs-qc:fastp-preprocess
raw reads (FASTQ) → cleaned reads
read-alignment:align-short-reads
cleaned reads → aligned BAM
variant-calling:call-variants-bcftools
aligned BAM → variants (VCF)
Data and compute
You run fragmentum on your own hardware. Here is where your inputs go and where the work happens, in plain terms, so you can judge it against your own data-handling rules.
- It runs on hardware you control. You install fragmentum on your own workstation or a private server. Each solution runs locally in its own conda environment, created with micromamba. There is no hidden cloud step.
- Your data files stay on your machine. Each solution reads the files you point it at and writes results back to the directories you choose, on the machine where you run the server; the files themselves are never uploaded. The agent is driven by a language model (Claude), so the text you type in the chat is sent to that provider, as with any AI assistant. Keep patient identifiers out of the chat.
- You decide what the agent may change. Search, inspection, and planning are read-only. Install, run, and deploy change the system, so they stay off until you enable them, and the agent asks before each one.
- Every result is attributable. Each solution declares its coordinates, arguments, citations, and license, so a result can be traced back to the exact code and method that produced it.
Research use only. fragmentum helps you discover and compose analysis software. It supports expert work and is not a medical device. It does not provide a diagnosis or any clinical recommendation. Every result depends on the underlying solutions and their inputs, and must be reviewed and confirmed by qualified people using independent evidence before it informs any decision about a person's care.