§0.0 ABSTRACT

Scientific software is broken.
Agents will fix it.

Fragmentum builds the infrastructure layer where any AI agent can discover, install, run, author, and publish reproducible scientific software, across any language and any platform. Born in research. Engineered for production.

Read §1 · The Protocol → See it run ↓

For investors, customers, and collaborators

The crisis is computational. And it is expensive.

Every year, scientists lose nearly half their working hours to dependency conflicts, environment failures, and DevOps that has nothing to do with their actual research¹. The United States alone burns through an estimated twenty-eight billion dollars in irreproducible preclinical research². The science is usually sound; the software simply cannot be re-run. Onboarding a single new computational tool into an enterprise pipeline takes six to eight weeks of engineering effort, every single time.

40% Researcher time lost to setup, conflicts, environment failures

$28B U.S. annual cost irreproducible computational research

6–8 Weeks per integration to onboard one tool into a pipeline

album + album-mcp

Album wraps any tool (Python, R, Java, C++) into a versioned, environment-isolated solution file. Solutions live in Git-hosted catalogs that you own. Album-mcp turns those catalogs into a substrate for agents: sixteen tools, full lifecycle, two-factor authentication on anything destructive.

       ┌──────────────────────────────────────────────────────┐
       │   AI AGENT       Claude, Cursor, Windsurf, custom    │
       └────────────────────────┬─────────────────────────────┘
                                │  MCP protocol
                                ▼
       ┌──────────────────────────────────────────────────────┐
       │   album-mcp        16 tools (5 gated by 2FA)         │
       │   discover, install, run, author, deploy             │
       └────────────────────────┬─────────────────────────────┘
                                │  lifecycle calls
                                ▼
       ┌──────────────────────────────────────────────────────┐
       │   album core       solutions, conda env, lifecycle   │
       │   versioned, reproducible, language-agnostic         │
       └────────────────────────┬─────────────────────────────┘
                                │  git fetch / deploy
                                ▼
       ┌──────────────────────────────────────────────────────┐
       │   Git-hosted catalogs        GitHub, GitLab, local   │
       │   copick, CellCanvas, DL4MicEverywhere               │
       └──────────────────────────────────────────────────────┘

FIG. 1 The fragmentum stack. Four layers: agent, protocol, framework, catalog.

What follows.

A complete account of what we have built, where it has been used, and why we believe the next decade of scientific computing will be agentic. Each chapter stands alone; together they make the case.

§	Chapter	Subject	Read
§1	The Protocol	16 MCP tools. Discovery, lifecycle, authoring, publishing, management. The full surface that turns album into agent infrastructure.	→ open
§2	Demonstration	An interactive transcript of an AI agent denoising and segmenting a microscopy dataset using album-mcp. Type, and see.	→ run
§3	Literature	Five publications across Nature Methods, Nature Communications, Nature Protocols, and arXiv. Where album is already used.	→ read
§4	Investigators	Two co-founders. Berlin. Sebastian Proft & Jan Philipp Albrecht. Roles still to be drawn.	→ meet
§5	Manifesto	Why the infrastructure layer stays open at the core. Why catalogs are decentralized. Why the agent is the user.	→ read

Album is already deployed.

A short ledger of where the framework appears in the published record and in production catalogs. Full citations in §3.

4 Publications Nat. Comms · Nat. Protocols · arXiv

10+ Public catalogs copick · CellCanvas · DL4MicEverywhere · &c.

1,200 Kaggle teams CZII CryoET · Chan Zuckerberg Initiative

$75K In prizes on the album-adjacent stack

The web gave us an internet of documents. MCP is giving us an internet of tools. Whoever owns the catalog of scientific tools, owns the rails. From §5, Manifesto

Talk to us.

If you run scientific computing in a lab, a research institute, or an enterprise R&D group, and you spend more time on environment failures than on science: we want to hear from you.

If you invest in deep-tech infrastructure and you see what agentic AI is doing to the substrate of computational research: we want to hear from you.

If you produce scientific tools and you want them discoverable, runnable, and citable across labs and companies: we want to hear from you.

fragmentum.ai@gmail.com

Continue to §1 → Skip to §5 manifesto ↪

The Cost of Reproducible Research. Survey data on developer time spent on environment setup vs. research, aggregated from multiple computational-biology workforce studies, 2022–2024. ↩
Freedman, Cockburn, Simcoe (2015). The Economics of Reproducibility in Preclinical Research. PLOS Biology 13(6). The $28B figure is the U.S. portion of irreproducible preclinical work that has a computational origin. DOI · ↩