CLIO
A context layer for scientific AI agents and workflows.
Context Layer for Input/Output, the architecture I created for IOWarp. Its shared contracts let Core, Kit, Agent, Relay, and Coder give scientific AI agents structured, queryable context, with each component built by its own team.
-
Who it serves
- Scientists and research software engineers
- Any MCP-capable agent: Claude Code, Cursor, Gemini CLI, VS Code
- Local models on one GPU, cloud APIs, or the ALCF inference gateway
-
CLIO agent-facing components
-
MCP servers, skills, and workflow plugins for HDF5, Slurm, ParaView, Pandas, NetCDF, FITS, Zarr, ArXiv, and more, installed selectively into agent clients.
Development team: Isa Muradli, Shazzadul Islam with Jaime Cernuda
-
The science agent: a Desktop and Web workspace over experts, scripts, live data, and local-first memory, with plugins and linked studies.
Day-to-day development led by Jaime Cernuda
-
The coding agent for the code science runs on. Typed tools, bounded workers, research loops, and sealed receipts.
Directly created and maintained by Anthony Kougkas
CLIO Relay and GACT TUI
Submit work to a remote cluster from a desktop tool and drive any compliant agent backend from a terminal or workspace.
Day-to-day development led by Jaime Cernuda
-
-
CLIO Core
The tiered context store and runtime
- Context Assimilation Engine Ingests 14+ formats
- Context Transfer Engine The Hermes lineage: placement across tiers
- Context Exploration Engine Query and retrieval
- Chimaera The runtime
Day-to-day engineering led by Luke Logan with the GRC core team
-
What it runs on
- GPU memory
- DRAM
- NVMe
- Parallel file systems
- Tape and cloud
Built on Ares at Illinois Tech, NSF DeltaAI, and ALCF Polaris and Aurora.
-
Research direction and architecture
Creator and architect of CLIO. Sets the research direction, component boundaries, shared contracts, integration, partnerships, and releases across the ecosystem.
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.
CLIO gives agents what databases give applications: 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 program and platform: the runtime, the engines, and the collaboration among GRC, The HDF Group, and the University of Utah. CLIO is the context architecture I created for it. I set its research direction, define the components and their contracts, coordinate integration across the ecosystem, and answer for what the system promises as a whole.
What shipped
CLIO is one architecture with technical leadership distributed across the team. I work across its layers while component leads and contributors carry the implementation, research, and releases.
CLIO Core is the tiered context store, with engines that assimilate, transfer, and explore scientific data on the Chimaera runtime. Luke Logan leads its day-to-day engineering. He built the runtime under Hermes as a student, wrote MegaMmap, and finished his PhD in 2025.
CLIO Agent is the science agent, a desktop and web workspace where a scientist works with data, tools, and papers alongside it. CLIO Relay lets a desktop tool submit work to a remote cluster and bring back logs, artifacts, and provenance, and the GACT TUI drives any compliant agent backend from a terminal. Jaime Cernuda leads all three. He built HStream and Hades, defended in 2025, and thinks in pipelines.
CLIO Kit collects the scientific tools and Model Context Protocol servers that agents call, from HDF5 and Slurm to ParaView and the National Data Platform. The students Isa Muradli and Shazzadul Islam build it, and Jaime and I contribute across its interfaces.
CLIO Coder is the component I build and maintain myself: a coding agent for the software science runs on, documented at coder.iowarp.ai, and my hands-on testbed for typed tools, bounded delegation, local models, and provenance.
CLIO Author, a team-built multi-agent harness for the scientific-paper lifecycle, CLIO Web Search, and the Agent Marketplace extend the same public ecosystem.
The loop that is coming
Two more components close the circle. CLIO Researcher is an unattended research loop for systems code and configuration: a person writes the directive, a planner model proposes one falsifiable change, CLIO Coder makes it in an isolated checkout, and a grader written as code decides whether the gain holds on a holdout split the planner never sees. A model never grades, and only a person merges the result. CLIO Bench is the record every run lands in, from Researcher and from the evaluations of every other component.
The first campaigns, in September 2026, tuned HDF5 chunk sizes and promoted one change each. Researcher is in public alpha, and Bench is unreleased.
What carried forward
CLIO inherits its working parts from the systems before it. The transfer engine in Core descends from Hermes. The task pipeline was reimplemented from DTIO in 2025. ChronoLog brought its MCP server in-tree with version 3.0 in May 2026, so a CLIO agent can record and replay time-ordered activity through Kit.
CLIO in the Genesis Mission
In July 2026 the Department of Energy selected two first-round Genesis Mission projects that build on CLIO. PORTUS, led by Nathan Tallent at PNNL, builds workflow performance models that guide coding agents as they port and tune scientific workflows on the Genesis Mission Platform; CLIO supplies the context and data services those agents use. SPOTTER-AI, led by Bogdan Nicolae at Argonne with Oak Ridge and HPE, captures provenance across Genesis workflows to trace and attribute threats to data and results; GRC contributes CLIO Core, Kit, Agent, and Coder. Both are Phase I projects that began in August 2026.
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 ships on its own rhythm. The shared contracts are what keep them one architecture.
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 first-class user.
- Getting the shared record schemas (clio-schemas) to carry the contracts I keep describing in prose.
- Building a shared regression corpus so “the agent got better” becomes a measurable claim.
Code, releases, papers, and records
- documentation
iowarp.ai platform hub
Overview of the CLIO stack, product portals, SC26 program, and documentation.
- release
CLIO Core v2.4.1, November 11, 2026
The tiered context and data runtime at core.iowarp.ai. Day-to-day engineering led by Luke Logan with the GRC core team.
- release
CLIO Kit v2.12.3, November 11, 2026
A scientific meta-marketplace of MCP servers, skills, and workflow plugins at toolkit.iowarp.ai. Developed by the Kit team; on PyPI as clio-kit.
- release
CLIO Agent v0.9.6.6, November 12, 2026
The science agent at clio.iowarp.ai, in beta, with Desktop and Web interfaces, plugins, and linked studies. Day-to-day development led by Jaime Cernuda; on PyPI as clio-agent.
- release
CLIO Coder v0.7.0, November 9, 2026
The coding agent I directly create and maintain, at coder.iowarp.ai, with model choice, managed peers, research campaigns, and sealed receipts.
- release
CLIO Relay v1.6.8
Remote execution from a desktop tool, with logs, artifacts, and provenance. Day-to-day development led by Jaime Cernuda.
- code
IOWarp organization on GitHub
Source repositories, issues, and contributor history across Core, Kit, Agent, Coder, Author, Researcher, Relay, and GACT.
- documentation
The CLIO platform on iowarp.ai
How CLIO Agent, Coder, Kit, and Core fit together.
- documentation
CLIO documentation
Install guides and first commands for each component.
- documentation
CLIO Core concepts
Tags and blobs, storage tiers and scores, persistence levels, pools, and the module chain.
- documentation
CLIO Kit catalogue
The scientific tools, skills, and workflows the kit publishes.
- documentation
CLIO Agent documentation
An introduction for researchers, from a first question to a result they can inspect and share.
Releases
-
CLIO Agent v0.9.6.6
lets the agent work on attached folders of any size where they already are, reads files with unusual names from connected sources, and fixes install problems on Linux and Windows.
-
CLIO Kit v2.12.3
adds offline MCP server environment preparation, self-healing server environments, the public National Data Platform catalog endpoint, and sandboxed Windows web-tool launches.
-
CLIO Core v2.4.1
stops the Unix-socket client receive thread from spinning a CPU core while idle, following the v2.4.0 replication and failover release.
-
CLIO Coder v0.7.0
adds the data tool for HDF5, NetCDF, and Parquet, sealed reproduce checks judged against a tolerance, a per-project claims ledger, and the research and arena commands.
People
- Luke Logan CLIO Core
- Jaime Cernuda CLIO Agent, CLIO Relay, GACT TUI
- Isa Muradli CLIO Kit
- Shazzadul Islam CLIO Kit
- Rajni Pawar CLIO Kit; Acropolis
- Eneko Gonzalez Research software engineering
- Xian-He Sun Long-standing collaborator
- Nathan Tallent PORTUS PI
- Bogdan Nicolae SPOTTER-AI PI
Related notes
- 2 min
CLIO in two Genesis Mission projects
Two first-round DOE Genesis Mission projects, PORTUS with Pacific Northwest and SPOTTER-AI with Argonne, build on CLIO. What each asks of it, and why context and provenance matter when agents act on shared infrastructure.
- 4 min
A study that keeps its evidence
CLIO Agent 0.9.5.1 links a study's questions, hypotheses, experiments, observations, and conclusions to the tool calls that produced them, so a scientific conclusion stays attached to its evidence the way a CLIO Coder receipt stays attached to its run.
- 3 min
CLIO Core 2.3 and 2.4: when a tier fills or a node dies
Four CLIO Core releases between October 9 and November 2, 2026 added a device-paged GPU vector and Intel GPU support, then taught the context store to recover a full tier and keep data and directories readable when a node goes down.
- 6 min
CLIO as the entry point
CLIO should become the single place a developer starts with IOWarp, built as an application SDK with deployment profiles rather than an umbrella package or a gateway service. This is the direction for 2027; the API and the scope are still open.
- 5 min
Context is part of the system
A capable model still fails when the system hands it the wrong slice of reality. CLIO's answer is to treat context the way a database treats state, and to seal a receipt for every action.
- 6 min
Run CLIO Coder against a model on your own GPU
Install CLIO Coder from npm, point it at a local llama.cpp or LM Studio server, run one headless turn, and read the receipt it seals.
- 8 min
Context is the new compute
The scarce resource for scientific AI is not FLOPS or bytes but assembled, trustworthy context, and the storage we built as tanks has to become a refinery. From the AI4Sys keynote at HPDC 2025.
- 8 min
From storage optimization to context orchestration
Twenty years of I/O research optimized bytes per second for programs. Agents are a different client, and each system in our lineage contributes one primitive to serving them, as argued in a seminar at Argonne.