Importing plugin packages
Understand which package components Mellow can use and what needs separate setup.
In this topic
Mellow can import compatible Claude-format plugin bundles and turn their supported pieces into Mellow skills, commands, schedules, and service connections. The import screen shows what a bundle contains before you install it. A recognizable format does not mean every extension declared by that bundle can execute in Mellow.
Inspect and install
Open the plugin catalog from Skills or Tools & MCP, or use the GitHub import action. Enter the repository, review the discovered plugins, and select the components you need. Review dependencies selected alongside the main plugin; a skill may refer to another skill in the same bundle.
If configuration is requested, complete the native form. Sensitive values belong in its secure fields rather than in chat or copied markdown. After installation, read the result summary for skipped items, missing configuration, and connections requiring sign-in.
For a first test, invoke one imported skill with a small task. Verify that any required service is connected before concluding that an installation problem is a model problem.
Where imported pieces appear
| Bundle item | Mellow destination | Follow-up |
|---|---|---|
| Skill instructions | Skills library | Review the instructions and referenced files |
| Commands | Chat slash-command list | Check required arguments |
| Scheduled-agent definitions | Schedules | Set a cadence and explicitly enable the schedule |
| HTTP or OAuth MCP definitions | Tools & MCP services | Finish authentication and tool discovery |
| Supporting documents and assets | References attached to the imported skill | Read when the task needs them |
| Unsupported declarations | Import/detail report | Do not rely on them at runtime |
Imported schedules start disabled when they need a cadence. A service row can exist before its credentials are configured. Neither state should be described as a fully operational plugin.
Configuration and local files
The importer records a stable plugin identity so related resources can be managed together. Settings declared through userConfig are collected by a configuration sheet. Non-secret settings and secret values use separate storage; secrets go through Keychain.
Compatible variable references include the plugin root, plugin data directory, project directory, declared user settings, and permitted environment values. These references must resolve to real resources for the imported service to start. A directory used for plugin data is different from the read-only imported package cache.
The repository importer reads supported plugin manifests and marketplace layouts. Check repository spelling, default branch, and manifest presence if no plugin is discovered. Do not assume an arbitrary repository containing markdown is an installable bundle.
Bundle format reference
| Artifact | Imported role |
|---|---|
skills/<name>/SKILL.md | Skill instructions |
Skill-local references, assets, and templates | Files attached to that skill |
agents/<name>.md | Schedule definition, with cadence setup when needed |
commands/<name>.md | Slash command |
.mcp.json | Supported MCP service definitions |
| Bundle-level reference documents | Context for imported skills |
The variable expander supports ${CLAUDE_PLUGIN_ROOT} for the package cache, ${CLAUDE_PLUGIN_DATA} for persistent plugin data, ${CLAUDE_PROJECT_DIR} for the current project context, and ${user_config.KEY} for declared configuration. Environment expansion is constrained to permitted or declared values. Secret configuration should be resolved through the secure storage path rather than copied into a manifest.
Updates and removal
Use the installed plugin's detail page to inspect components, configuration, and available updates. Updating replaces the bundle's managed resources instead of intentionally creating a second copy. Review the new component list because an update can add requirements.
Uninstall removes resources managed by that imported plugin. Review the operation before continuing if the plugin has schedules, service connections, or local data you need. Removing a local service grant does not necessarily delete the account or all tokens on the external provider.
Compatibility boundaries
Hooks, language-server declarations, output styles, channel declarations, executable exports, and other format-specific extensions may be recorded without being executed. Read the detail page's unsupported-items notice. A helper script attached as reference material is not automatically run merely because it was included.
Stdio MCP definitions depend on the supported local or sandbox execution path and its configuration. An imported disabled entry needs setup; do not simply enable an unknown command to clear a warning.
Troubleshooting
Repository cannot be read: confirm access and inspect GitHub rate-limit or authentication errors. Add a token only through the supported secure configuration flow.
Skill imports but a tool is missing: check the plugin's service connection, required configuration, and discovery result.
Scheduled work never runs: open the imported schedule, set its time, choose a usable model, and enable it.
A declared extension has no effect: inspect the compatibility notice rather than repeatedly reinstalling.
Update changed behavior: compare the imported instructions and component list, then test one narrow workflow before restoring unattended use.
See Skills, MCP services, and Schedules for the runtime controls behind each imported component.
Continue exploring · Connect your workspaceConnecting other clients →Choose the right integration path for model clients, tools and agent execution.