Getting Started

CLIO Kit is a meta-marketplace for scientific MCP servers, workflow skills, plugins, agent definitions, hooks and external marketplace collections. Choose individual components or a plugin that bundles what your workflow needs. Skill and agent collections can also be installed separately. External packages and marketplaces remain under their maintainers' control. See components and plugins for the catalogue and installation model.

The shipped scientific servers are Python projects. The launcher also supports Node/TypeScript and Go descriptors with runtime-specific dependency locks. See Marketplace and Contributions for supported components, contribution commands, update behavior and tested runtime boundaries.

For client-specific commands, use Set up your agent. For the individual MCP, skill and plugin routes, follow Install components, then try the terminal tutorials.

Install from source

Use the same checkout for the launcher and marketplace. The commands below install the checked-out code; package releases are published separately.

git clone https://github.com/iowarp/clio-kit.git
cd clio-kit
uv tool install --force --reinstall --editable ".[verification]"
clio-kit mcp-servers

Install uv first. If the launcher is not on PATH, run uv tool update-shell and open a new shell. The editable launcher reads component sources from this checkout. For selective downloads without a checkout, see release installation. First server launches may download dependencies.

Portable skills: Codex and other compatible agents

clio-kit skill install --bundle clio-scientific-io --target /path/to/project/.agents/skills

For Codex, .agents/skills is a project skill directory; ~/.agents/skills is user-wide. Other agents use their own documented discovery path as --target. Omit --bundle for all bundled skills, including adapted Clio Coder skills and the dataset-report procedure. Configure required MCP servers separately; for example codex mcp add clio-hdf5 -- clio-kit mcp-server hdf5 registers HDF5. The scientific I/O workflow also requires ADIOS, Parquet and compression.

The optional npm skills CLI can also install these folders into agent projects. It does not configure MCP servers.

Claude Code native plugins

claude plugin marketplace add "$PWD"
claude plugin install clio-scientific-io@clio-kit
claude mcp list

This route requires Claude Code with plugin dependency support. Reload plugins or restart an existing session before using its new tools. Plugin registration alone does not establish that servers connect.

The repository's setup guide includes an actual compression round-trip check and troubleshooting steps.

Choose a workflow

Bundle	Work
`clio-hpc`	Software discovery, JARVIS execution and Slurm scheduling
`clio-performance`	I/O diagnosis, log investigation and session provenance
`clio-scientific-io`	Inspect and read HDF5, ADIOS BP5, Parquet and compressed files
`clio-analysis`	Tabular data, statistics, charts and ParaView
`clio-geoscience`	Geospatial, terrain and seismic analysis
`clio-research`	Literature, public datasets and operator catalogues

In Claude Code, a bundle installs its member servers and matching skills. Install a server alone with claude plugin install clio-hdf5@clio-kit, a workflow's procedures with clio-scientific-io-skills@clio-kit, or all procedures with clio-skills@clio-kit. clio-agents@clio-kit provides a workflow planner and evidence reviewer.

Agent integrations

Install a workflow for your client

Install a workflow's skills and MCP configuration together into a working project. A released launcher uses its component catalogue; inside a CLIO Kit checkout, the command uses local packages instead:

clio-kit plugin install clio-scientific-io --client codex --project /path/to/project

Choose codex, opencode, cursor, antigravity, vscode or claude-code. Use --root /path/to/clio-kit when running elsewhere, and --dry-run to inspect the plan. New local package folders are discovered without a separate sync. The command resolves selected dependencies, obtains complete skill folders and merges stdio MCP entries into the client's project configuration. It does not change global settings or install a native client plugin. Install the launcher first. Keep source checkouts available when using --root; released package scripts use the component cache. Codex and Antigravity share .agents/skills; installing the same skill there once is enough, and replacing it affects both clients.

Existing unrelated configuration is preserved. Conflicting names stop installation; review before using --replace. Changed configurations get a recovery backup; TOML comments remain in that backup. OpenCode JSONC and remote/host-specific MCP options require manual configuration. Maintained packages and indexed GitHub, Git URL, Git subdirectory and npm packages use the same installer, including federated entries.

Agents and hooks: Codex, Claude Code and OpenCode have adapters for the scientific agents and Dataset Report hooks. Other native components need their own supported adapter; installation refuses to silently omit them. Use --components-only only when you deliberately want skills and MCPs without the other components. Review Codex hooks through /hooks before use. See the client support table.

Reload your client and verify skills and MCP connections. Codex project settings require a trusted project. For Antigravity CLI, first open the project using agy --new-project. Installation alone does not prove a model used the workflow.

Individual component configuration

Client format references: Codex plugins, OpenCode MCPs and skills, Cursor MCPs, and Antigravity MCPs.

MCP servers provide tools; skills provide procedures for using them. Configure both for a complete workflow. The scientific I/O workflow needs hdf5, adios, parquet, and compression, plus its three skills.

Client	Register MCP servers	Project skill target
Codex CLI / IDE extension	`codex mcp add`; shared `~/.codex/config.toml`	`.agents/skills`
Claude Code	Native marketplace above, or `claude mcp add --scope project`	Bundle-managed, or `.claude/skills`
Cursor	`.cursor/mcp.json` with `mcpServers`	`.cursor/skills`
OpenCode	`opencode.json` with `mcp` (`type: local`, command array)	`.agents/skills`
VS Code / GitHub Copilot	`.vscode/mcp.json` with `servers`	`.github/skills`
Antigravity	MCP settings → raw config; workspace `.agents/mcp_config.json`	`.agents/skills`
Claude Desktop	Developer settings → `claude_desktop_config.json` with `mcpServers`	Local MCP setup does not install skills

For example, in Codex:

codex mcp add clio-hdf5 -- clio-kit mcp-server hdf5
codex mcp add clio-adios -- clio-kit mcp-server adios
codex mcp add clio-parquet -- clio-kit mcp-server parquet
codex mcp add clio-compression -- clio-kit mcp-server compression
clio-kit skill install --bundle clio-scientific-io --target .agents/skills
codex mcp list

For Cursor, Antigravity, or Claude Desktop, merge these entries into the configuration file listed above:

{
  "mcpServers": {
    "clio-hdf5": { "command": "clio-kit", "args": ["mcp-server", "hdf5"] },
    "clio-adios": { "command": "clio-kit", "args": ["mcp-server", "adios"] },
    "clio-parquet": { "command": "clio-kit", "args": ["mcp-server", "parquet"] },
    "clio-compression": { "command": "clio-kit", "args": ["mcp-server", "compression"] }
  }
}

For VS Code / GitHub Copilot, use this complete .vscode/mcp.json structure:

{
  "servers": {
    "clio-hdf5": { "type": "stdio", "command": "clio-kit", "args": ["mcp-server", "hdf5"] },
    "clio-adios": { "type": "stdio", "command": "clio-kit", "args": ["mcp-server", "adios"] },
    "clio-parquet": { "type": "stdio", "command": "clio-kit", "args": ["mcp-server", "parquet"] },
    "clio-compression": { "type": "stdio", "command": "clio-kit", "args": ["mcp-server", "compression"] }
  }
}

Run MCP: List Servers to start the servers, then use Copilot's Agent mode. The Codex extension uses Codex's own configuration, even when running in VS Code.

Install skills using the appropriate target from the table:

# Cursor
clio-kit skill install --bundle clio-scientific-io --target .cursor/skills
# VS Code / GitHub Copilot
clio-kit skill install --bundle clio-scientific-io --target .github/skills
# Antigravity
clio-kit skill install --bundle clio-scientific-io --target .agents/skills

Current Cursor, VS Code, and Antigravity also discover .agents/skills; reuse a shared installation there instead of making duplicate copies. Antigravity also supports the older .agent/skills directory. Its current global locations are ~/.gemini/config/mcp_config.json for servers and ~/.gemini/config/skills for skills. Use the IDE's MCP Servers → Manage MCP Servers → View raw config to locate the configuration for your installed version. For Antigravity CLI, start the first session from your project with agy --new-project so its local skills and MCP configuration are loaded. Reopen it with agy --project <project-name-or-id>.

The .claude-plugin export remains available for Claude's native marketplace. Kit's project installer adapts shared packages for the selected client; see component support for agents and hooks. Claude Desktop's MCP configuration installs only servers. For another local stdio client, use command clio-kit with arguments ["mcp-server", "SERVER_NAME"] in its documented schema. Remote-only clients need a separately deployed MCP endpoint.

Install the launcher in the environment where your client starts its MCP processes, including SSH, WSL, or a container. If clio-kit is not on the client's PATH, use its absolute executable path. Reload or restart the client, inspect its discovered tools and skills, and continue with verification below.

Client references: Codex MCP and skills, Cursor MCP and skills, VS Code MCP and skills, Antigravity MCP and skills, and Claude Desktop MCP.

Verify the workflow

clio-kit doctor --server hdf5 --connect checks prerequisites and a real MCP connection. Follow it with a representative tool call on known data and inspect the result. HPC workflows require site software and scheduler access; ParaView and ChronoLog require compatible native backends. A successful connection does not establish that those dependencies are available or that a scientific result is correct.

The marketplace guide describes fresh-wheel installation tests, real client installation, Node/Go fixture execution, scientific data checks and remaining acceptance boundaries.

Repository structure

clio-kit/
├── .claude-plugin/marketplace.json
├── mcp-servers/            # Independent scientific server projects
├── plugins/                # Workflow bundles, aggregate skills and agents
├── skills/                 # Workflow and imported skill collections
├── community/entries/      # External plugin and marketplace entries
├── src/clio_kit/            # Launcher, authoring, discovery and cache management
├── tests/fixtures/         # Runtime acceptance fixtures
├── docs/                   # Shared human documentation
├── website/        # Website configuration and UI
└── AGENTS.md               # Agent contribution entry point

Server environments are isolated and identified by source and lock contents. Python projects use uv.lock, Node projects use package-lock.json, and Go projects use module versions and checksums. Released component artifacts download separately from the launcher. A cold installation requires network access; see offline preparation.

Contribute and get help

Start with Authoring components for plugin, MCP, skill and hook recipes, or the documentation index for other guides. Read the contributor guide for local skills, hosted servers and indexed contributions. Use clio-kit plugin init, plugin validate and plugin submit for plugin authoring. The marketplace guide explains what each command validates and publishes.

CLIO Kit is developed by the Gnosis Research Center at Illinois Institute of Technology as part of IOWarp, with National Science Foundation support. The repository is licensed under BSD-3-Clause.

Report issues on GitHub or join the community discussion.

Install from source​

Portable skills: Codex and other compatible agents​

Claude Code native plugins​

Choose a workflow​

Agent integrations​

Install a workflow for your client​

Individual component configuration​

Verify the workflow​

Repository structure​

Contribute and get help​