CLIO in 2026: from agent to marketplace
CLIO is the agent layer of IOWarp, GRC's platform for managing scientific data and context for AI. In February 2026 it had two parts, a storage runtime and a tool kit. By November 13 it had six, each with its own release schedule, and together they had published more than 300 tagged releases on GitHub this year. This is a tour of what each part does and how it got there.
The base: CLIO Core and CLIO Kit
CLIO Core provides the storage and runtime services underneath everything else: it keeps data and context on distributed, tiered storage close to where agents and applications need it. Core shipped 24 releases between version 0.6.5 on February 23 and version 2.4.1 on November 11. Version 2.3.0, released October 9, brought the GPU runtime up on Intel GPUs, raising the number of passing tests on Argonne's Aurora system from 104 to 254. The same release served HDF5 reads from the CLIO cache tier, cut the latency of a cross-node collective operation by 3.1 times, and fixed seven crash-consistency bugs found across five fault scenarios. Luke Logan leads Core development.
CLIO Kit connects agents to scientific and HPC tools through Model Context Protocol (MCP) servers. It went from version 2.0.1 in February to 2.12.7 in November, about 60 releases in all. Jaime Cernuda and Isa Muradli are its main contributors.
An agent for science: CLIO Agent and its desktop
CLIO Agent made its first 2026 release, version 0.3.1, on April 27, and reached version 0.9.7 on November 13 after nearly 90 releases. Version 0.9.0 in July added three ways to control what the agent does: a planning mode in which the agent may write only its plan until a person approves it, project hooks that run a lab's own scripts before or after each action, and sessions that loop, pursue a stated goal or run on a schedule. All three go through one permission system, where a deny rule always beats an allow.
Version 0.9.7 builds web and scholarly search into the agent. Its literature tools query OpenAlex, Crossref, arXiv, Europe PMC, OSTI.gov, DataCite and OpenCitations, and ask before the first query to each source. Deep research now delivers its report with full references, and the agent can see the plots and page renders its own scripts open. Long conversions and builds run until they finish instead of hitting a fixed time limit.
People reach the agent through GACT TUI, a terminal and desktop front end that can drive any compliant agent back end. It went from version 0.2.1 in April to 0.13.0 on November 13, about 60 releases. Version 0.13 adds Spanish, Japanese and Greek to the interface, marked as machine translations until native speakers review them, along with a guided tour for new users. Cernuda leads the agent and its front end.
A marketplace of science agents
The CLIO Agent Marketplace holds the agents and shared capabilities that CLIO Agent loads. Its first release, version 0.5.2 on June 17, shipped agents that analyze EarthScope GNSS station data through the National Data Platform; they passed all seven tasks of the team's test matrix with three different open-weight models.
Version 0.7.0, released October 28, replaced that first layout with agent classes and capability plugins that agents can combine. It introduced CLIO Data, which inspects, checks and converts HDF5, netCDF-4, ADIOS2 and Parquet data and keeps its conversion checks with its procedures. Its deep-research guidance records verified claims, rejected claims and open gaps separately, so a report shows which evidence still needs checking. Version 0.8.0 followed on November 13, the twelfth marketplace release of the year.
CLIO Coder
CLIO Coder is an open-source coding agent for scientific software and everyday engineering. The experimental first release on April 24 already ran worker processes for batch tasks, kept a record of every run and gated tool access through three safety modes. About 60 releases later, version 0.8.0 on November 13 targets people who run a coding agent beside colleagues, other agents and scripts, such as researchers on a shared login node. Two sessions can share a folder safely, the sandbox hides credentials and token files from contained commands, and headless runs end with one machine-readable result for automation. A per-project store keeps results and datasets outside the working tree, and diagnostic reports are shown to the user before anything leaves the machine. A new guide covers installing it on HPC login nodes. Anthony Kougkas leads its development.
Around the stack
Two releases outside the CLIO repositories support the same work. Jarvis-CD 1.8, released August 6 and maintained by Logan, ships execution profiles for scientific applications, services and I/O interceptors that the maintainers keep current, the same ones used in a cross-cluster evaluation campaign. It also reconciles Slurm jobs that never reached the Jarvis runtime against live and accounting state.
FITS-HDF5-VOL, released May 6 and written by Muradli, lets HDF5 programs read FITS astronomy files without converting them. It is a read-only HDF5 Virtual Object Layer connector, so tools such as h5py and h5dump open an unmodified FITS file and read its pixels, tables and headers. It is validated against HDF5 1.14.x and 2.1.x with 162 tests, including the astropy test corpus and files from NRAO.
CLIO will be shown at the HDF5 birds-of-a-feather session at SC26 on November 19; see GRC at SC26.