CLIO

What CLIO is

CLIO stands for Context Layer for Input/Output. It is named for the Greek muse of history, because the whole point is that an agent working on science should be able to remember, cite, and account for what it did.

Here is the one-line version I keep coming back to: CLIO gives agents what databases give applications, which is structured, reliable, queryable state. A language model on its own has a prompt and a context window. A scientist has terabytes on a parallel file system, a Slurm queue, instrument metadata, a decade of provenance, and a set of tools with side effects. CLIO is the layer in between.

How it relates to IOWarp

IOWarp is the NSF-funded platform: a distributed runtime, a Context Transfer Engine for moving data across storage tiers, a Context Assimilation Engine for reshaping it, and the institutional collaboration around all of that. Xian-He Sun is the award’s principal investigator; I am a co-PI.

CLIO is the brand and the architecture I lead on top of that platform. Through IOWarp, I am CLIO’s overall PI and architect. That means I decide what the components are, what contracts sit between them, and what each one is allowed to promise. My team owns and builds the components. I build one of them myself.

The components, and who builds them

• CLIO Core is the tiered context and data store, with the Chimaera runtime underneath. Luke Logan leads it with the GRC core team. Released and maintained; v2.2.1 is current.
• CLIO Kit is a suite of Model Context Protocol servers for scientific computing: HDF5, Slurm, ParaView, Pandas, NetCDF, FITS, Zarr, ArXiv, compression, hardware, and more. Jaime Cernuda, Isa Muradli, Soham Sonar, Shazzadul Islam Mun, Aum Sathwara, and others built it, with contributions from me. Released and maintained; v2.10.5 is current and on PyPI.
• CLIO Agent is the science agent: a planner over experts and tools, with a local-first memory layer. Jaime Cernuda leads it, with contributions from me. Beta, v0.9.1.1.
• CLIO Relay and GACT TUI let a desktop tool submit work to a remote cluster and drive any compliant agent backend from a terminal. Jaime Cernuda’s work, v1.6.8 and v0.9.9.
• Clio Coder is the coding agent for the people who maintain scientific software. I created it and I maintain it. It has its own story.
• CLIO Researcher and CLIO Bench are mine and still private: an auto-researcher for I/O systems and a benchmark suite for agentic behavior. Alpha at best.
• One more private, team-built component is on the way; the team will name it when it opens.

I list the names because “my team” is not an abstraction. These are the people whose commit histories you will find in the repositories.

The bet

The bet is that context deserves systems infrastructure of its own, and that the interesting engineering is in the contracts between components rather than inside any one agent loop. What can be discovered. What can be transferred without copying. What is durable across sessions and machines. What requires a person to approve. What leaves a receipt.

Every component answers those questions differently, and that is fine. What I refuse to do is let a diagram stand in for a release. Core and Kit are released and maintained. Agent is beta. Coder is experimental and ships quickly. Researcher and Bench are private. When a component’s maturity changes, this page changes.

What I am pushing on now

• Making the receipt a first-class object across components, not only inside Clio Coder.
• Keeping local models the design target everywhere, so a lab with one GPU is a real user and not an afterthought.
• Getting the shared record schemas (clio-schemas) to carry the contracts I keep describing in prose.
• Turning CLIO Bench into a public regression corpus so “the agent got better” becomes a measurable claim.