Skip to main content
All tutorials

Contribute 6 min read

Build, test and contribute a plugin

Package a skill with an MCP, test it in Claude Code, Codex, Clio Coder and OpenCode, and share it through the community catalogue.

Written for
CLIO Kit 2.11
Works in
Claude Code · Codex · Clio Coder · OpenCode
Basis
Recorded sessions · 7 Oct 2026
Claude Code loads the contributed skill and calls its Pandas MCP.
Claude Code loads the contributed skill and calls its Pandas MCP.Claude Code · 2.1.293

This walkthrough packages a small check-runtime-table skill with the Pandas MCP. It shows the same package used as a native Claude Code plugin and as portable components in Codex, Clio Coder and OpenCode. All four sessions were run for this guide with the versions listed in Explore an HDF5 file; they produced the same checked group means from the same input.

1. Start with the tested package​

Install the CLIO Kit launcher. Create a new directory and download lab-checks.zip into it:

mkdir contribution-demo
cd contribution-demo
unzip lab-checks.zip

The archive contains only these authoring files:

.claude-plugin/marketplace.json        # local test index
plugins/lab-checks/
├── .claude-plugin/plugin.json        # name, description, author, version
├── .mcp.json                         # starts the clio-pandas MCP
└── skills/check-runtime-table/
    ├── SKILL.md                      # when and how to check the table
    └── evals.md                      # expected behavior on example inputs

To scaffold your own package from scratch, use clio-kit plugin init plugins/my-plugin, then replace its sample content. The download above is the finished package used in the two sessions below, not an installed marketplace contribution. Its local test index lets Codex resolve it before you open a PR.

2. Read and adapt the skill​

The included skill requires configuration, trial and runtime_s, checks coverage and missing values before grouping, retains the tool's output file, and asks for one independently checked mean. It preserves input files and stops when a required column or MCP is missing.

Keep that scope narrow. Set an accurate trigger in the YAML description, document the required MCP, and provide an expected result in evals.md. Replace the manifest's tutorial author and description before contributing your own package.

The MCP configuration is a normal stdio entry:

{
  "mcpServers": {
    "clio-pandas": {
      "command": "clio-kit",
      "args": ["mcp-server", "pandas"]
    }
  }
}

A skill-only contribution can omit the MCP entry. Hook and agent packages require host-specific checks; use the authoring guide for those formats.

3. Validate the package​

clio-kit plugin validate plugins/lab-checks
claude plugin validate plugins/lab-checks --strict

Both validators passed for the downloaded package. They check structure, not whether its procedure produces a scientifically correct result.

Create a fresh test project outside contribution-demo for each client you try. Put a copy of runs.csv in it and keep an untouched copy for comparison.

4. Run it in your client​

Use /absolute/path/to/contribution-demo for the folder you unzipped.

Load the package as a native plugin for one session:

claude --plugin-dir /absolute/path/to/contribution-demo/plugins/lab-checks

Enable the plugin's MCP server, then ask:

Use the check-runtime-table skill from lab-checks to check runs.csv. Runtime values are in seconds. Read the skill and call its Pandas MCP, check coverage and missingness, compute group means and verify one independently. Preserve the original. Save runtime-check.md and report output paths. Work only in this temporary project, no network or delegation.

Claude loads lab-checks:check-runtime-table; the namespaced prefix shows the skill comes from the plugin. It then calls the plugin's Pandas MCP:

Claude Code loads lab-checks, hashes runs.csv, and calls the plugin's clio-pandas profile and groupby_operations tools

Claude Code reports full coverage, no missing values, means of 12.5 and 9.5 seconds, an unchanged input hash and the output paths

5. Check the outputs​

In each project:

cat runtime-check.md
ls *.csv

All four sessions found 8 rows, no missing values, and means of 12.5 seconds for baseline and 9.5 seconds for tuned. The grouped file is runs_grouped.csv in the Codex, Clio Coder and OpenCode runs and runtime_means_by_configuration.csv in the Claude Code run; open the path your agent reports. We checked each grouped CSV and confirmed every original runs.csv was unchanged. The report should state units and coverage without claiming a real performance improvement from this fixture.

The included missing-column scenario is an additional test to run when adapting the package; the screenshots above demonstrate the complete-input case.

6. Choose a contribution route​

Maintained here: put your reviewed package in plugins/<name>/ in the CLIO Kit checkout and open a repository PR. Website builds and CI discover valid folders and update the catalogue. Skills, agents and hooks have corresponding package folders; see contribution routes.

Maintained in your own repository: follow the community steps below. CLIO Kit indexes one entry; your implementation, dependencies and releases stay with you.

7. Prepare a community entry​

Publish the contents of plugins/lab-checks/ in your own repository, with .claude-plugin/plugin.json at its root. Include the .mcp.json and skills/ directory, a license and installation instructions. Replace the tutorial author with your name and rerun the validation and usage checks above.

Use your real GitHub repository in place of your-org/lab-checks:

clio-kit plugin submit plugins/lab-checks \
  --repo your-org/lab-checks --output lab-checks.toml
cat lab-checks.toml

The entry contains your manifest's name, description and maintainer, followed by:

[source]
type = "github"
repo = "your-org/lab-checks"

This command validates the local package and writes a file. It does not publish your code, check that the remote matches it, or open a PR. Confirm the published repository includes the files you tested. The entry filename must match its name; community names cannot start with the reserved clio- prefix.

If you keep the package inside a larger repository instead, replace the source table with its actual subdirectory:

[source]
type = "git-subdir"
url = "https://github.com/your-org/your-repo.git"
path = "plugins/lab-checks"

See the community source formats for npm plugin packages, Git URLs and revision pinning. An npm MCP executable alone needs a plugin wrapper with .mcp.json before it can use this route.

8. Submit the entry for review​

In your CLIO Kit fork, add the reviewed file as community/entries/lab-checks.toml and open a PR against iowarp/clio-kit. Only the entry belongs in that PR; the plugin stays in your repository. Catalogue generation runs automatically during the build and CI.

For a plugin at the root of a GitHub repository, you can alternatively let the CLI create the submission. Sign in with gh auth login, then run:

clio-kit plugin submit plugins/lab-checks \
  --repo your-org/lab-checks --open-pr

This command creates or uses your GitHub fork, pushes a new branch and opens a public PR. It generates a fresh GitHub entry; use the manual route if you edited the source table or added a revision pin. Neither route merges the contribution.

Include the installation route, client versions, inputs, output checks and limitations in your PR. Keep model credentials and local session directories out of the contribution. Review covers the entry and a minimal installation/usage check; indexing does not certify every external command or future release.

9. Install the accepted contribution​

After your entry is merged and available in the catalogue, Claude Code users register the marketplace once, then install your plugin:

claude plugin marketplace add iowarp/clio-kit
claude plugin install lab-checks@clio-kit

If the marketplace is already registered, run claude plugin marketplace update clio-kit before installing. Reload plugins or restart Claude Code, then repeat the runtime-table request from step 4 and check the outputs from step 6. This example still requires the CLIO Kit launcher for its Pandas MCP.

When you publish an update, bump the plugin manifest version. Existing users run:

claude plugin marketplace update clio-kit
claude plugin update lab-checks@clio-kit

Reload the plugin after updating. These are native Claude Code installation commands. Kit also installs indexed packages through the shared project installer; agents and hooks need an adapter for the selected host. For Codex, follow the local component installation in step 5 using the tutorial's local index, or provide your own client-specific instructions.

Contribute a whole marketplace​

If you maintain a collection, submit its index instead of each plugin:

clio-kit plugin submit /path/to/your-marketplace \
  --repo your-org/your-marketplace --kind marketplace \
  --output your-marketplace.toml

It needs .claude-plugin/marketplace.json with metadata.description. Submit the generated entry under the filename reported by the command. Maintainers refresh the federated catalogue to include its plugins; the implementations remain in your repository. See marketplace federation for refresh behavior and conflict handling.

The recorded sessions above demonstrate local plugin usage. The community steps describe publishing your own contribution; this tutorial does not claim that your-org/lab-checks exists or that the example has been accepted into CLIO Kit.