Documentation lives with the code.

IOWarp is the team at the Gnosis Research Center that builds CLIO. CLIO, the Context Layer for I/O, is the open platform it builds: the agentic frameworks CLIO Agent and CLIO Coder, the CLIO Kit tools they share, the CLIO Core context runtime, and the GPU, storage, application, and research components around them. It matters because organizations can run AI agents against scientific data and keep a record their people can check.

Choose the task first. Install CLIO Agent to work with scientific data, or CLIO Coder to work on scientific software; they are different products for different audiences. Install CLIO Kit to give any MCP-compatible agent scientific tools, and CLIO Core for the context runtime and its APIs. Each installs on its own and has its own prerequisites.

Install the standalone runtime with pip install iowarp-core. Check the PyPI files for current CPython versions, operating systems, and architectures. GPU runtimes come as separate CUDA, ROCm, and SYCL wheels or source builds. MPI and some format adapters need a source build. Linux wheels need libfuse3 to mount, Windows needs WinFsp, and macOS mounting requires a source build with macFUSE. Keep Agent’s tested Core dependency when installing Agent. Published wheel compatibility on PyPI ↗ Mount on macOS ↗

For the Desktop app bundled with its agent, run curl -fsSL https://clio.iowarp.ai/desktop.sh | bash on Apple Silicon macOS or Linux, or irm https://clio.iowarp.ai/desktop.ps1 | iex in PowerShell on Windows, or download a package at clio.iowarp.ai. For the local service and web interface only, run curl -fsSL https://clio.iowarp.ai/install.sh | bash on Linux or macOS, or irm https://clio.iowarp.ai/install.ps1 | iex on Windows, then open it with clio web or start it headless with clio start. The installer downloads the prebuilt release runtime and provisions Agent’s tested Core and Kit dependencies. Agent installation guide ↗ Release installation reference ↗ Connect a model ↗

The shell installer needs no root and supplies its own Node.js runtime on Linux and macOS. PowerShell and CMD installers are available for Windows, where native interactive use remains best effort. An npm install uses your own supported Node.js runtime. Connect a model that supports tool calling, then run clio-coder doctor to check the installation and connections. Coder documentation and installation ↗

Install the standalone CLIO Kit launcher with uv tool install clio-kit. Run clio-kit mcp-servers to list its servers, then clio-kit mcp-server hdf5 to fetch and start the HDF5 component. Components download on demand, and server dependencies stay isolated and locked. Follow the client guide to register a server or install a workflow plugin. Selective installation ↗ Client setup and component support ↗

Agent keeps Core in its tested compatible release family and pins Kit for its marketplace plugins. Keep that combination when installing Agent. Standalone Core and Kit have separate release and upgrade paths; a newer standalone package does not establish Agent compatibility. Agent’s tested dependency declarations ↗ Agent installer and its Kit pin ↗ Standalone Kit installation ↗

It depends on the component. CLIO Kit has tools for formats such as HDF5, Parquet, ADIOS2, SAC, and geospatial data. CLIO Core has its own assimilation adapters, including HDF5 in source builds. Check the specific server or adapter before relying on a format in a workflow.

Yes. CLIO Core ships clio_cee for bundling, querying, retrieving, and destroying contexts, and clio_cte_core_ext for tags, blobs, and search. Its Python guide explains runtime attachment and indexing requirements. Core Python API ↗

Library packs and content plugins provide reusable instructions and procedures. Sandboxed extensions load after approval of their declared capabilities, and playbooks describe fleet contracts. A semantic context index is optional; background indexing is off by default, and scanned PDF recognition and raw video embedding are unsupported. The research and arena commands use the checks you declare in the project. The desktop app remains alpha. Research and arena guide ↗ Released behavior and limits ↗

The current flagships, scientific context runtime, and supporting tools are open source, and IOWarp will launch the full stack at SC26 in Chicago. GPU and operating-system backends have build and hardware prerequisites. Archived BaM and GPU–NVMe bridge prototypes remain in older tagged source and are absent from the current Core release. Current Core release ↗ Archived prototypes ↗

Open an issue in the repository of the component you used, with the version, environment, input format, expected and observed results, and a small reproducible example if you can. Zulip is open for discussion.